HOME
QUICK SEARCH
AUTHORS
KEYWORDS
HELP
BUGSPharmaSUG CN 2026 Papers - Index
70 papers. Source: PharmaSUG (Pharmaceutical Software Users Group) conference proceedings.
| # | Title× | Authors× | Abstract× | Keywords× | Extras | Pages |
|---|---|---|---|---|---|---|
| 1 | A Comprehensive R Shiny Application for Pharmacokinetic Analysis | Keyi Zhang | This paper describes an interactive R Shiny application that streamlines PK statistical analysis across common clinical pharmacology study types. The tool supports plasma concentration profiling, non -compartmental analysis (NCA), dose proportionality asses sment, food effect evaluation, drug-drug interaction (DDI) analysis, bioequivalence (BE) testing, special populations exploration, human ADME mass balance evaluation, and BE sample size estimation. For each module, the app generates summary tables of key PK parameters and produces customizab le visualizations (e.g., concentration time curves, boxplots, forest plots, and dose proportionality scatter plots). Users can upload their own datasets in standard formats, select analysis parameters interactively, and download results in publication ready formats. The application is built with modular R code to facilitate maintenance and extension. By integrating multiple PK workflows into a single, user friendly interface, this tool reduces repetitive coding effort, enhances reproducibility, and supports informed decision making during drug development | forest plot, sample size, geomean, pk parameter, sample, confidence interval, scatter plot, box plot, compartmental analysis, degrees of freedom, proc mixed, standard deviation, proc means, dose, time variable, regression analysis | source | 17 |
| 2 | A Desktop Application for Automating PDT Management, SAS Execution, and Report Assembly | Mingjie Fu | Statistical programming teams track the development and validation of every table, listing, figure, and analysis dataset in a Program Development Tracking (PDT) workbook stored on a shared network drive. As studies grow, this single EXCEL file becomes a bottleneck: programmers scroll through hundreds of rows to find their own assignments, edits collide when several people open the workbook at once, and there is no automatic way to tell whether an output is still current with the raw data it depends on. This paper describes PDT Manager, a Windows desktop application built with Python and PySide6 that wraps the existing PDT workbook in a faster, safer interface. The tool reads the workbook into a local cache, lets programmers filter to their own tasks in one click, and writes edits back to the original EXCEL file through a controlled save path that detects concurrent modification before overwriting. Beyond tracking, it integrates two operational steps that previously required leaving the workbook: it submits SAS programs to a remote server through SASpy and parses the returned log for errors, warnings, and PROC COMPARE results; and it assembles individual output files into a single combined report with a generated cover page and table of contents. A freshness indicator on each row compares output and source file timestamps and propagates a "stale" state along a user-defined dependency graph, so a change in an upstream SDTM dataset visibly flags every downstream output that needs to be rerun. The application has been used on production studies at the author's organization. This paper focuses on the design decisions that made the tool safe to use against a live shared workbook and the techniques that other teams can reuse for similar tracking-and-automation problems. Keywords: statistical programming, Python, PySide6, SAS automation, workflow tracking |
proc step, proc compare, cdisc | source | 5 |
| 3 | A Dual-Engine Architecture for MedDRA Auto-Encoding: Integrating Rule-Based Logic with LLM Semantic Retrieval in R Shiny | Qingqing Wang, Yaxiong Suo |
Adverse event (AE) coding using the MedDRA dictionary is essential for clinical trial safety analysis but remains labor-intensive due to the dictionary's extensive scope (>60,000 preferred terms). Manual coding suffers from inefficiency and inconsistency, while existing automated approaches often lack generalizability or interpretability. To address this, w e developed a dual -engine auto -coding system through the R Shiny framework, integrating deterministic rule-based matching with large language model (LLM)-powered semantic retrieval. The rule engine prioritizes project-specific historical mappings and sponsor-defined conventions, delivering high-precision exact and pattern matching. For unmatched terms, a LLM scans the MedDRA preferred terms, enabling semantic similarity matching that returns ranked candidate codes with quantitative scores. The Shiny application supports end -to-end workflows including data ingestion, coding processing, interactive visualization, and comprehensive version comparison . Internal validation demonstrated that semantic retrieval achieved 90% accuracy, substantially reducing the need of manual dictionary lookups. This dual-engine architecture effectively balances precision and efficiency, leveraging deterministic rule - based consistency alongside LLM semantic generalization. The framework provides a reproducible, transparent, and extensible solution for AE coding standardization, with applicability to broader clinical data transformation tasks | clinical trial, adverse events, quality control | source | 6 |
| 4 | A Dual-Harness Workflow to Wrap AI into Observable, Reusable Clinical Table Delivery: From SAP Intent to Shell-Aligned Execution | Wenjun Zhu | Statistical programming teams working in a single therapeutic area repeat a familiar pipeline for every study: read the Statistical Analysis Plan (SAP), pick and adapt table shells, bind them to Analysis Data Model (ADaM) variables, run analyses, and produce accurate results. Modern coding agents (Claude Code, GitHub Copilot, Codex, etc.) can also cope with these routines and are multipotent -they generate code, execute it, edit files, and increasingly assemble deliverables end-to-end inside a single session. What they do not, by themselves, provide is the domain validation gates, audit trails, and institutional memory that regulated table, listing, and figure (TLF) production requires. This paper treats that gap as a control problem rather than a prompting problem. A large language model is a powerful but stochastic actuator; a regulated pipeline demands deterministic, explainable outcomes. The workflow presented here reconciles the two by wrapping the model in a harness-a feedback controller that supplies the reference signal (the sponsor's table shell, treated as the executable spec of the analysis), the sensors (deterministic validation gates), the comparator (shell -aligned checks and human confirm/reject decisions), and the recorded state (an append-only store in which every confirm, reject, and patch becomes structured data). Within this loop, evaluation separates into two arms - semantic SAP-to-shell selection and deterministic shell-aligned execution-connected by a typed spec[] handoff placed exactly where statistical programmers have always handed work off to one another, so the boundary that made human review accountable makes agent-assisted review accountable as well. Two properties of this design travel beyond the specific system | data cleaning, audit trail, phase i, phase ii, statistical programmer, data science, cdisc | source | 14 |
| 5 | A Hybrid Framework for SAS Code Quality Review: Combining Deterministic Static Rules with Large Language Model Semantic Analysis | Meng Zhang | Clinical-trialSASprogramsforSDTM,ADaM,andTLF deliveryareusuallyreviewedmanuallyagainst GoodProgrammingPractice(GPP) guidance,study specifications,and companySOPs.Manualreviewis effective,butitis difficulttoapplyconsistently acrosseveryprogram,everyreviewer,andeverydelivery cycle.This paperpresentsatool(hereafter referredtoasSAS CodeReviewer),alocally deployable reviewframeworkthatcombines deterministic staticcheckswithanoptionalsemantic reviewperformed byaLarge LanguageModel(LLM).ThesystemfirstrunsaPythonruleengine ofSAS-specificchecks coveringsyntax,naming,comments,macrohygiene,maintainability,safety, hardcodechecks,and selectedlogic hazards.Thestatic findingsarethenpassedintoastructuredpromptsothat theselected LLMcanconfirm,refine,andsupplementtherule-basedfindingsrather thanreasoning fromscratch.The toolsupportsbothLLM APIcallsandlocally deployedmodels,anduserscanswitchbetween providersat runtimewithout code changes.Thearchitecture separatesaVue /Monaco frontend, aFastAPI backend, andaunifiedprovider adapter;deterministicandLLM findingsaremergedintoasingle,deduplicated reviewreport.Thesamesetofreviewrules is alsopackagedasaSkill,sothatitcanbeinvokeddirectly insidecompatibleAIassistants | attrib statement, run | source | 11 |
| 6 | A Practical Framework for Applying Generative AI in Clinical SAS Programming: A Human-in-the-Loop Approach for Chinese Clinical Studies | Siyu Yu | Objective: This paper proposes a practical framework for applying generative artificial intelligence in clinical SAS programming. It does not position AI as a replacement for clinical programmers, but embeds it in a controlled, specification-driven, human-reviewed workflow for protocol interpretation, specification review, SAS development, log checking, and output validation. The framework defines a risk-scaled release gate: the evidence needed before an AI-assisted draft is review-ready. Background: Clinical SAS programmers are required to work with multiple sources of information, including protocols, statistical analysis plans, SDTM and ADaM specifications, TFL shells, annotated CRFs, define.xml, study data review guides, project macros, and study-specific folder structures. In Chinese clinical study environments, additional challenges often arise from Chinese CRF content, Chinese text values, UTF-8 encoding, SAS Viya compatibility, special characters, and the need to maintain consistency between Chinese source documents and English submission deliverables. Traditional programming workflows rely heavily on programmer experience and manual review, which may lead to repetitive work, inconsistent coding styles, environment-specific issues, and time-consuming log troubleshooting. Generative AI provides new opportunities to support these tasks, but its use must be standardized and carefully reviewed to avoid logic errors, traceability gaps, and quality risks. Methods: A four-component AI-assisted clinical programming framework was developed. First, common programming tasks were classified into practical categories, such as specification interpretation, SDTM and ADaM derivation, TFL programming, batch execution, log review, and submission document support | pharma programming, clinical sas, data cleaning, sas programming, independent qc, cdisc | source | 12 |
| 7 | A Shiny-Based Tool for SAS and R Batch Execution and Log Review | Meng Zhang, Nan Zhao |
Statistical programmers in clinical studies routinely submit large batches of SAS programs and then inspect every resulting log for errors, warnings, and a broader set of quality-relevant diagnostic messages such as character truncation, uninitialized variables, unresolved macro references, and zero-observation outputs. As R has been increasingly adopted for derivation of datasets and production of tables, listings, and figures, the same review effort now extends to a second language. Most existing batch utilities, however, are single-language tools - built exclusively for either SAS or R - leaving programmers to maintain two separate toolchains and two sets of log-review rules. This paper presents a lightweight Shiny-based toolkit (BatchTool) that unifies batch execution and log scanning across both SAS (sas.exe and SAS Enterprise Guide) and R (any installed version, with automatic renv pinning detection). The tool is deployed as two folders only: a shared root that lives at a fixed, team-agreed location and holds YAML rule files and the execution engine; and a tiny per-project front-end folder that is copied into each study and double-clicked to launch. We describe the architecture, core technical modules, serial batch route, and in-browser review workflow. Using an example folder derived from public clinical-programming examples, we demonstrate the folder layout, configuration points, rule examples, and run/log-check workflow so readers can reproduce the design or adapt individual components | sas enterprise guide, sas log | source | 12 |
| 8 | A Standardized Programming Framework for Daily Dose Dataset Construction from Multi-Source Clinical Data | Xianjun Li, Chengye Lu |
Glucocorticoids (GCs), such as prednisone, are widely used in clinical trials and are often central to evaluating steroid-sparing treatment effects. However, deriving reliable daily and cumulative GC exposure is technically complex because dose information may originate from multiple sources, including study - defined taper schedules, concomitant systemic GC medications, and rescue or escape treatments. This paper presents a standardized ADaM-based programming framework for constructing a daily GC dose dataset by aligning multi-source GC exposure on a subject-day timeline, converting all eligible GC doses to prednisone equivalents, and retaining separate dose components before aggregation. The resulting dataset supports consistent derivation of daily and cumulative exposure endpoints, estimand- aligned intercurrent event handling, and downstream analyses such as cumulative dose through Week 52 or adjusted annual cumulative dose. The framework improves traceability, reproducibility, reviewability, and cross-study consistency in GC exposure analysis for regulatory submission deliverables | intercurrent event, proc freq, concomitant medications, bootstrap, imputation, sample, adam dataset, clinical trial, composite strategy, study day, treatment effect, treatment period, while-on-treatment, cdisc | source | 10 |
| 9 | A Traceable and Reproducible Framework for Oncology Efficacy Data Review Using IMWG Criteria: A BCMA Case Study | Yun Jiang, Yingying Deng, Manman Lu, Longhui Li |
Oncology efficacy review often requires complex algorithmic assessment across multiple clinical data sources. In BCMA studies, efficacy review based on International Myeloma Working Group criteria requires integration of laboratory, bone marrow, imaging, and clinical assessment data. Because response outcomes are derived through multiple rule-based steps, manual review can be time-consuming and difficult to standardize. This paper presents a traceable and reproducible framework for IMWG-based efficacy data review. The framework was developed in R within Positron and supports SDTM-to-ADaM derivation through modular, parameter-level steps. Each step generates intermediate outputs, issue reports, and links to the final response result. A dashboard supports study-level monitoring, subject-level review, parameter-level traceability, and issue investigation, while an AI assistant helps users understand algorithm logic, derivation decisions, and response assessment evidence. The framework converts complex response assessment into a transparent and reproducible workflow. Parameter-level outputs allow programmers to trace how individual assessments contribute to endpoint results, while automated issue reports and interactive review tools improve consistency, efficiency, and interpretability. This approach provides a practical model for applying automation and AI-assisted review to clinical analyses with complex endpoint derivations | data cleaning, analysis set, cdisc | source | 6 |
| 10 | ADICE: A Proposed ADaM Dataset Structure for Intercurrent Events in the Estimand Framework | Lijie Zhu | The ICH E9(R1) addendum introduced the estimand framework, making the precise definition and handling of intercurrent events (ICEs) a central requirement in clinical trial design and analysis. Despite this, no standardized ADaM dataset structure exists to capture, classify, and trace ICEs in a consistent, submission-ready way. This paper proposes ADICE, an ADaM-compliant Occurrence Data Structure dataset designed to fill this gap. ADICE consolidates all trial-level ICE records into a single subject-level dataset with full traceability to source SDTM domains. It supports multiple estimand strategies - composite, hypothetical, treatment policy, and while-on-treatment - through explicit strategy variables and ICE type flags, enabling downstream efficacy analysis datasets to derive estimand-aligned analysis flags from a single authoritative source. This paper describes the ADICE dataset structure, key variables, data flow, CDISC standards alignment, and integration with efficacy endpoints. Dummy data examples illustrate single and multiple ICE scenarios, including cases where different estimand strategies apply to the same subject under a defined priority hierarchy | composite strategy, while-on-treatment, treatment policy strategy, adam dataset, flag variable, intercurrent event, efficacy summary, hypothetical strategy, imputation, clinical trial, sdtm domain, principal stratum, cdisc | source | 11 |
| 11 | AI in Regulatory Submissions: Health Authority Expectations and Emerging Opportunities for Data Standards and Submission | He Liu | Every major regulatory agency-FDA, EMA, PMDA, and NMPA-has now taken a public position on AI in drug development. This paper examines two landmark regulatory developments-FDA's January 2025 draft guidance on AI use in drug development and the joint FDA-EMA ten guiding principles published in January 2026-and their implications for AI use in regulatory submissions. The FDA draft guidance establishes a seven-step credibility framework for AI that produces information or data to support regulatory decision-making, and states that it does not address AI used for operational efficiencies that do not impact patient safety, drug quality, or the reliability of results. The joint FDA-EMA ten guiding principles apply to all AI use in drug development, with validation and oversight proportionate to risk | clinical research, regulatory submission, data cleaning | source | 7 |
| 12 | AI intelligent document translation: The deep integration of high-fidelity typesetting and large model knowledge bases | Zuocheng Xiang | Regulatory submissions, legal contracts, and clinical documentation in the life-sciences industry increasingly require multilingual versions that preserve both semantic accuracy and typographic fidelity. Traditional machine-translation (MT) pipelines discard document structure, resulting in formatting-loss rates exceeding 40%. Computer-assisted translation (CAT) tools rely on 100% manual post-editing, creating an obvious throughput ceiling. This paper presents an AI-Agent-driven translation architecture that unites (1) a four-stage structural pipeline-deep parsing, context-aware translation, structure-level re-rendering, and dual-layer quality assurance; (2) LLMWiki, a translation-native knowledge base that stores structured terminology, transformation rules, style guidelines, and correction cases; and (3) a closed-loop feedback mechanism in which every post-editing correction automatically enriches the knowledge base, creating a self-reinforcing quality flywheel. In production validation across pharmaceutical regulatory packages, the system reduced formatting errors from over 40% to below 8%, raised cross-document terminology consistency from approximately 60% to 95%, and lowered human post-editing revision rates from 40% to 15%. Delivery cycles compressed from weeks to hours for ten-thousand-page projects, while marginal cost trends toward zero as the knowledge base matures. | quality assurance, data cleaning | source | 8 |
| 13 | AI-Assisted Automated Development of the P21C Validator Report Comparison Tool | Huanhuan Sun | During project execution, the Pinnacle 21 Community (P21C) validator report needs to be run multiple times. Traditional manual comparison of old and new reports leads to repetitive work, high error rates, and inefficiency, while making it difficult to effectively track issue closure status. This is especially problematic in later project stages when version differences are minimal, yet every detail still requires line-by-line verification, resulting in resource waste. To address this business pain point, we have developed an automated solution designed to intelligently manage the lifecycle of P21C report comparisons. This solution leverages Generative AI-assisted development model to accomplish code development, logic debugging, and rapid feature iteration. Its core technical approach is centered on structured field recognition and intelligent difference matching, allowing it to adapt to custom column name variations across different projects. The system achieves automatic discrepancy identification between two report versions, identification of new issues, classifies closed-loop statuses, and intelligently migrates historical comments, enabling a one-click comparison of new and old P21 reports. This solution significantly reduces the manual verification workload and lowers error rates, proving especially valuable for the rapid confirmation of stable versions in later project phases. Concurrently, it assists Statistical Programmers (SPs) in summarizing data issues for Data Managers (DMs) to perform data cleaning, thereby fostering a more effective and closed-loop workflow between them. This initiative strongly advances the evolution of clinical trial data quality management towards automation, intelligence, and standardization, aligning perfectly with the long-term industry trends of cost reduction, efficiency improvement, and proactive quality management | data cleaning, clinical data | source | 6 |
| 14 | AI-Assisted Code Generation for Structured Table: A Structured Filtering Logic Extraction Approach | Qing Ma | AI is reshaping statistical programming for structured tables. In these settings (e.g., some efficacy outputs), shells are standardized and variation stems from endpoint-specific selection. We present an approach that externalizes and formalizes endpoint-level filtering logic. A small set of reference programs anchors consistency; iterative AI identifies, normalizes, and organizes rules, then reuses them to generate batch programs aligned with predefined shells. Logic expansion is separated from validation: AI standardizes and consolidates rules, while experts perform targeted checks for consistency and clinical correctness. Traceability is supported by a rule registry with provenance, diffs across iterations, and mappings from rules to code and table cells. When data are structured and patterns stable, the method emphasizes logic reuse, high completenes s, and efficient review. Efficacy serves as an example in the paper and the approach applies to any standardized table stream, turning implicit logic into reusable, auditable artifacts that improve transparency, scalability, and consistency | clinical trial, phase iii | source | 6 |
| 15 | AI-Assisted QC for Clinical Trial Tables: A Scalable Approach for Heterogeneous Templates and Distributed Data. | Shuchun (Coral) Wu | Quality control (QC) programming for clinical trial tables is a time-consuming task that becomes significantly more complex when the study involves a large number of outputs, multiple data sources with varying structures, and rule sets that are spread across different documents. These conditions are common in safety analyses and cannot easily be handled by a single automated process. This paper describes a practical workflow that uses AI assistance to address this complexity. Rather than attempting to generate fully complete QC programs automatically, the approach produces programs that are approximately 70-90% complete and clearly marks the remaining gaps for human follow-up. The workflow is organized into three layers - context structuring, controlled generation, and structured review - and is designed so that each review cycle makes the next generation cycle more accurate | clinical trial, analysis set | source | 7 |
| 16 | AI-Augmented SDTM Workflow: A Methodological Framework | Yuqi Chen, Muqu Lv, Zhuozhao Zhan |
Inclinical trialdatasubmissions,theStudyDataTabulation Model(SDTM)acceleratesdrugapproval throughstandardized structures.TraditionalSDTMgenerationistime-consuming,relyingheavilyon manualCaseReportForm(CRF)annotation,specificationwriting,anddatasetsprogramming.To addressthesechallenges,thispaperpresents anAI-augmentedworkflowframeworkthatautomatesthe end-to-endmappingprocess.Thisframework integrates programmaticexecution,LargeLanguage Model (LLM)automation,andhumanoversight:(1)Programmaticexecutionhandlesfoundational tasks, includingCRFreconstructionanddomainspecificationtemplateinitialization;(2)Buildingonthe structuredCRF andspecificationtemplate,LLMautomatesthemanual efforts:draftingannotations, completingvariableconversionrules,andgeneratinganddebugging executablecode;(3)Rigorous reviewmechanismsensurethe qualityofAI-generatedcontent,includingcollaborativeAI-human refinementandmandatoryhumanapproval.Comparedtoconventionalprocesses,this framework significantlyreducesmanual effortwhileensuringCDISCcomplianceandmethodological consistency acrossprojects. Keywords: SDTM, LLM, workflow automation, CRF annotation, specification, code generation | sdtm programming, data cleaning, cdisc | source | 16 |
| 17 | AI-Native Real-Time SDTM Mapping and Exploration Agent | Hao Chen, Xuejie Zhou, Haomin Yang, Qianwang Wang |
This paper introduces SATA, the SDTM Automated Transformation Agent, an AI -native, real-time solution for SDTM mapping, transformation, and exploration. SATA is designed as an integrated all -in-one web application that enables data ingestion, transformation, validation, and exploration across diverse inputs, including raw clinical datasets, annotated CRFs, and metadata repositories. The platform provides real - time analysis, traceability, and on-demand partial reruns, allowing users to dynamically refine mappings and immediately assess results. This reduces feedback cycles and improves development efficiency. The system combines intelligent SDTM specification prediction, automated SAS/R code generation, sandbox-based execution, iterative self-testing, and human-in-the-loop validation. Designed as a "digital employee," SATA integrates domain knowledge, clinical metadata, and generative AI to enable end-to- end automation from EDC data to SDTM. Beyond transformation, SATA also supports SDTM -based dashboards, TFL generation, and ad hoc question answering, accelerating data -driven decision-making processes. Initial implementation has demonstrated substantial efficiency gains, reducing SDTM mapping timelines from days to hours or even minutes while improving consistency, scalability, and transparency. This paper presents SATA's system design, architecture, development progress, use cases, and future directions, including AI-native mapping and continuous model optimization. SATA establishes a scalable foundation for next-generation clinical data analysis automation |
sdtm programming, clinical data, raw data, data cleaning, sdtm domain, data science, data transformation, clinical trial, phase iii, production run, response variable, sample, cdisc | source | 13 |
| 18 | AI-Powered R Shiny Tool for Automated Footnote Comparison Across TLF Packages and Shells | Yuan Sun | InthepreparationofTLF(Tables,Listings,andFigures)packages,acriticalstepiscomparingtitles and footnotesbetween packagesand correspondingshells.Usually,footnoteverification hasbeenperformed manually-aprocessthatisboth time-consumingandpronetohumanerror.Toenhancetheefficiency and accuracyofthis qualitycontrol process,wedeveloped anAI-powered RShiny toolthatautomatically captures andcomparesfootnotesacrosspackages andshells. Thispaper presentsthedesign,functionality, andimplementationofthisvalidation tool,demonstrating howautomation cansignificantlyimprovetheTLFreviewworkflowinstatisticalprogramming. Additionally, we discussthe benefits ofAI-drivendevelopmentinaccelerating innovationand improving programmerproductivity instatisticalprogrammingenvironments | adverse events, data validation rules | source | 11 |
| 19 | AI-Powered Spec Verification for Data Review Outputs Using ALS Cross-Checks | Keting Lv | We developed an AI-driven tool to streamline and standardize specification verification for clinical data review deliverables. Manual creation and review of output specifications (e.g., Exception Listings, Reconciliation Listings, and Metric Reports) are prone to typographical errors, inconsistent wording, and cross-document mismatches; they also require substantial effort to check and maintain. Our workflow automatically ingests key inputs such as the ALS and output specification documents, performs structured comparisons across fields and rules, and flags discrepancies including redundancies, typos, missing/extra requirements, inconsistent terminology, and logic mismatches. The tool generates issues in a batch and traceable manner, producing review-ready findings that can be used for peer review and vendor oversight. By shifting verification from ad hoc manual checks to systematic, repeatable AI-assisted controls, this approach improves consistency, reduces cycle time, and helps minimize specification-to- output defects before downstream programming and QC | excel, data cleaning | source | 6 |
| 20 | An Integrated, AI-Augmented R Shiny Platform for Oncology Clinical Trial Decision Intelligence: From CDISC Data to Go/No-Go in One Unified Hub | Zhiping Yan | Modern oncology drug development necessitates that cross-functional teams integrate safety, efficacy, pharmacokinetic/pharmacodynamic (PK/PD), and competitive benchmark evidence to make coherent and timely decisions, often within strict clinical timelines. Nevertheless, in most organizations, these analyses remain fragmented across isolated tools, manual scripts, and static reports, resulting in both operational inefficiencies and decision delays. We introduce the Oncology Clinical Trial Data-Driven Decision-Making System, an open-architecture R Shiny application consisting of over 40 modular analytical domains. The platform takes in CDISC - standard datasets (ADSL, ADAE, ADRS, ADTR, ADTTE) and user-uploaded PK exposure or benchmark data, and then unifies them via a reactive data layer that supplies: (1) a Bayesian decision hub that implements Beta-Binomial Go/No-Go analysis, Bayesian Logistic Regression CRM with Escalation With Overdose Control (EWOC), BOIN/mTPI dose-escalation simulation, Simon two-stage design, and joint efficacy-toxicity Phase II design; (2) an AI Signal Intelligence Engine that actively scans six evidence domains and ranks actionable alerts in real-time; (3) a machine-learning safety strategy module that combines logistic regression patient risk scoring, K-means AE clustering, and sequential CUSUM monitoring; (4) a competitive benchmark intelligence engine with Meta-Analytic Predictive (MAP) Prior elicitation; and (5) an AE co-occurrence knowledge graph. The architecture shows that modern statistical programming practices, such as reactive modular design, Bayesian computation, and lightweight ML, can transform a passive reporting tool into an active decision co-pilot. This paper details the system design, statistical methodology, implementation choices, and lessons learned, aiming to provide a reproducible blueprint for the broader biopharmaceutical programming community | phase ii, sample, data cleaning, sample size, phase i, posterior probability, audit trail, monte carlo, phase iii, dose escalation, dose, type i error, regression analysis, logistic regression, binomial distribution, logistic distribution, bar chart, clinical trial, dose modification, stage 1, adverse events, credible interval, data cutoff, dose finding, clinical research, efficacy summary, insurance risk score, stage 2, proc phreg, forest plot, cdisc | source | 28 |
| 21 | An Interactive Shiny Application for Streamlined ADaM Specification Integration and Compliant Define-XML v2.1 Generation | Dongxu Liu | Define-XML v2.1 is the mandatory machine-readable metadata companion for clinical data submissions to regulatory agencies. Despite clear CDISC standards, manual production of Define -XML remains widespread, leading to reproducibility gaps, late-cycle defects, and submission risk. This paper presents an R-based Shiny application that automates the full Define-XML v2.1 lifecycle: metadata ingestion from structured EXCEL workbooks, automatic detection and dual support of both SDTM and ADaM standards, deterministic XML generation using the R xml2 package, optional XSLT-driven PDF rendering, and a comprehensive two-tier validation framework covering over 100 metadata quality rules and the full published CDISC DD/OD rule catalogue. The application is built entirely in R using shiny, bslib, xml2, xslt, readxl, dplyr, stringr, and purrr. It provides a multi-tab interactive UI for metadata browsing, validation, generation, and conformance checking, while preserving programmatic, headless execution capability for integration into submission pipelines. Key technical innovations include automatic SDTM/ADaM standard detection from workbook metadata, value-level metadata with flexible WHERE_CLAUSE and WHERE_VARIABLE/OPERATOR/VALUE syntaxes, ADaM AnalysisResultDisplays support, and structured validation findings tied to named CDISC rule IDs. The framework demonstrates how modern R tooling can replace fragile document -based workflows with a reproducible, auditable, and submission-ready metadata production pipeline |
data cleaning, excel, xls, cdisc | source | 11 |
| 22 | An Interpretable Machine Learning Framework for Exploratory Survival and Safety Analysis Using Random Survival Forests and SHAP | Kun Shao | Inclinical trials,standardizedCDISCAnalysisDataModel(ADaM) datasetscontainrichbaseline characteristicsandcomplex efficacydataresults.Traditionalsurvival models,particularly theCox proportionalhazardsmodel, remaincentraltoconfirmatoryanalyses.However,exploratoryclinicaltrial questionsofteninvolvehigh-dimensionalbaselinecovariates,non-linearrelationships,potential interactions,and astrongneedforclinical interpretability. This paperpresentsanADaM-compatible, interpretablemachinelearningframeworkforexploratorysurvivaland safetyanalysesusing Random SurvivalForests(RSF) andSHapleyAdditiveexPlanations (SHAP).Theframeworkincludesendpoint definition,featuregovernancetoreduceinformationleakage,dual-pathmodelingwithfull-featureRSFand Cox-Lasso-selectedRSF,SHAP-basedinterpretation,andtraditional statisticalbenchmarking.The workflowisdemonstratedusingdata fromaclinicaltrial,focusingonProgression-FreeSurvival (PFS)and targetedriskendpointsderivedfromstandardADSL,ADTTE,andADAEdatasets.Theproposed frameworkcomplementstraditional statisticalmethods bysupportinginterpretable hypothesisgeneration, riskprofiling, andexploratorysignaldetectionwithin familiarclinical trialdata structures | adverse events, clinical trial, feature selection, machine learning, data cleaning, cdisc | source | 17 |
| 23 | Applying AI LLM to Reduce Manual Translation Workload in E-Data Package Submission | Megan Chen | In 2020, NMPA issued the Guideline on the Submission of Clinical Trial Data, mandating foreign language database translation and setting minimum requirements for e-data package documents. Despite internal teams' efforts in developing in-house translation tools, much content still requires time-consuming manual translation and review. We explored AI + LLM to reduce this workload, aiming to evaluate its effectiveness, develop a quality scoring process. We selected some representative content for AI + LLM translation, with the AI + LLM workflow fine-tuned to internal and NMPA standards. Results showed 60% less translation time, 45% less review workload, 80% score overlap degree between quality scores and manual reviews, and full NMPA compliance. AI + LLM effectively reduces manual translation burden, with the quality score facilitating efficient review and optimizing e-data package submission | clinical trial, regulatory submission, clinical data | source | 7 |
| 24 | Automated Clinical Text Classification with Robust Machine Learning for Pharmaceutical Data Curation | Weineng Zhou | In clinical trial data management, table-of-contents (TOC) EXCEL files must be annotated with CDISC datasetidentifiers(e.g.,ADSL,ADAE,ADLB) beforedownstreamstatisticalanalysis.AtClinflash,labeling thousands of Chinese-language table titles acrossdozens ofstudies was performedmanually-aprocess requiring hours per study and producing inconsistent results between reviewers. This paper presents a lightweight Python tool that automates TOC dataset labeling using TF-IDF text vectorization (ngram_range=(1,2), max_features=5000) and multinomial logistic regression (lbfgs solver, max_iter=1000). The tool consists of two scripts: predict_dataset_cn.py, a one-click production tool that trains on all available labeled TOC EXCEL files in the working directory and predicts dataset labels for new studies via a tkinter file-selection dialog, outputting a _dataset_filled.xlsx file; and model.py, an evaluation script that performs stratified 80/20 train-test splitting, binary classification assessment (largest class vs. rest), and exports nine clinical metrics to evaluation_report.xlsx alongside four publication-quality plots (ROC, PR, calibration curves, confusion matrix). Trained on 1,849 rows aggregated from 14 clinical study TOCfiles spanning60 CDISC datasetlabels,theproduction modelhandles multi-classpredictiondirectly. On a held-out binary evaluation, the model achieves accuracy=0.8859, ROC-AUC=0.9349, sensitivity=0.6522, and specificity=0.9398. A standalone Windows executable (predict_dataset_cn.exe), built with PyInstaller, allows non-technical users to run the tool without installing Python or any dependencies. KEYWORDS: TF-IDF, Logistic Regression, Text Classification, CDISC, Table of Contents, Clinical Data Curation,Python,tkinter GUI,PyInstaller |
logistic regression, excel, logistic distribution, regression analysis, confusion matrix, training set, clinical trial, cdisc | source | 11 |
| 25 | Automating ADaM Specification with AI: Intelligent Generation, Risk Assessment, and Human-in-the-Loop Review | Shuang Gao, Yaohui Zhu, Tingting Zeng, Jinling Li, Wei Wang |
Authoring Analysis Data Model (ADaM) specifications is a knowledge-intensive activity in clinical statistical programming. A complete specification must align CDISC standards, study design, source data, statistical analysis plan requirements, protocol definitions, historical study practices, and internal programming conventions. Adapting existing materials to a new study still requires substantial manual effort, because many decisions depend on current-study evidence and expert interpretation. This paper presents a controlled AI approach to ADaM specification authoring. The system generates structured metadata candidates for human review, drawing on company standards, historical study references, and current-study documents. A read-only Ask Agent supports day-to-day ADaM knowledge queries. AI is used to assist drafting and surface evidence; final decisions remain with expert reviewers. The paper describes what the system produces, how reviewers interact with it, how candidates are evaluated, and what governance controls are in place | data science, clinical programming, cdisc | source | 11 |
| 26 | Automating SDTM Programming with a Human-in-the-Loop Multi-Agent System | Wenqiang Chu | This paper presents a multi-agent system for automating the SDTM programming lifecycle with an explicit human-in-the-loop control model. The system comprises three core agents with distinct but complementary responsibilities. The Coder Agent performs completeness checking, generates structured implementation plans, and produces executable SAS programs from SDTM mapping specifications. The Reviewer Agent evaluates the generated outputs, produces a structured review report, and returns findings for downstream routing rather than modifying code invisibly. The Docs-Researcher Agent handles standards-related knowledge queries from other agents by searching SDTMIG documents and the CDISC Library, returning grounded evidence snippets. These agents are coordinated by a deterministic orchestrator that advances execution and pauses only at explicit human decision gates. The current system supports question-answer clarification, plan review, validation handoff, review decision, audit manifests, and a lightweight web UI for monitoring and intervention. This architecture reduces manual effort, improves consistency, strengthens traceability, and turns SDTM standardization into a scalable AI-assisted workflow rather than a purely manual bottleneck | data cleaning, sdtm programming, clinical data, excel, cdisc | source | 7 |
| 27 | Breaking Silos via Recycled Metadata: TOC-Driven Automation for Standard Title and Footnote Generation | Tracy Cui | In clinical statistical programming and document development, fragmented file management and repetitive manual configuration of title and footnote (TLaFT) create significant information silos. These silos result in redundant metadata entry, inconsistent formatting and inefficient document iteration. This paper presents a practical R-Shiny automation solution that breaks down document silos by reusing TLaFT metadata. The application offers three core functions: 1. Metadata Recycling - Automatically retrieves TLaFT metadata from previous deliverables, even across compounds or studies. 2. Intelligent Matching - Automatically aligns standard title-footnote templates with key information derived from pre-defined TOC metadata. 3. Regulatory-Ready Output - Automatically generates accurate, standardized, and regulatory- compliant titles and footnotes for final TFL production | clinical programming, excel, clinical trial | source | 7 |
| 28 | Bridging gtsummary and reporter: Submission-Ready Clinical Outputs with derrick | Xiecheng Gu, Joe Zhu, Bill Huang |
Clinical reporting in R often divides naturally into two layers. Packages such as gtsummary, crane, cardinal, CARDS, and cardx are well suited to defining table content and statistical structure, whereas reporter is designed to control how a clinical report is laid out and written to a file. The practical difficulty is the repeated translation between these layers. Column labels, spanning headers, indentation, widths, pagination, titles, footnotes, and output-specific rules can otherwise require a separate setup for every table. derrick turns that translation into a reusable package interface. Its main function, gtsummary_reporter(), accepts either a gtsummary object or a report-ready data frame, interprets the available display metadata, applies shared clinical layout rules, and delegates file generation to reporter. A single call can produce RTF, TXT, DOCX, PDF, or HTML output while preserving the table definition established upstream. The package is the production extension of an earlier PharmaRUG derrick | Technical Paper 2 prototype: the original function demonstrated that the two ecosystems could be connected; derrick adds validation, documentation, compatibility handling, tests, and a versioned distribution model. This paper describes the design of the bridge, the three supported table-construction paths, the translation mechanism, the report controls provided by the package, and the engineering practices that make the workflow reusable across clinical reporting programs. Keywords: R, gtsummary, reporter, clinical reporting, RTF, TXT, DOCX, PDF, HTML 1. Clinical reporting is a two-layer problem Clinical tables must satisfy two different concerns. First, the program must define the analysis result: variables, populations, denominators, statistics, labels, row hierarchy, and the order in which results are presented |
data cleaning, tables | source | 12 |
| 29 | Chronic Graft-versus-Host Disease (cGVHD) efficacy criteria exploration and implementation challenges - based on a first line phase 3 case study | Ting Zhang | Chronicgraftversushostdisease(cGVHD) isanimmune-mediatedinflammatoryandfibrotic disorder characterizedbytissuedamageandmultipleorganinvolvement.Itistheleadingcauseofmorbidity,late non-relapsedmortality(NRM),andimpairedquality oflifeafteranallogeneic hematopoietic cell transplantation(HCT),affecting upto70% ofallallogeneic HCTrecipients.Duetoits widespreadand considerablediseaseburden,ithasacceleratedtheevolutionforinternationalconsensus response criteria. In2005,theNationalInstitutes ofHealth(NIH)cGVHDConsensusResponseCriteriaWorkingGroup initiallyestablishedtherecommendationsformeasures andresponseevaluationcriteria.And revisedby the2014WorkingGrouptoenhancereliability andpractical utilityinclinical trialstoaccelerate developmentandapprovalofnoveltherapeuticagents incGVHD.Per recentFDArecommendationwhen initiatingfirstlinephase3study, basedon2014NIHconsensus responsecriteria,it's suggestedtouse progressionfromnadir insteadofbaselineafterthefirstcompleteresponseorpartial responseachieved andprogressionconfirmationmayneedperspecificscenario.Based onexistingstandardandFDA recommendations,theexplorationandimplementation ofthenewefficacycriteriaembarkedonwhich providedhighlyvaluableexperienceforfutureclinicaltrialsdevelopmentnotonlyforcGVHDindication andalsoforother indicationregardingthechallengesand goodpracticeduringtheinitialization, standardization,andimplementation ofnewefficacycriteria. Thispaper willpresent howitworksforbothcurrentstandardandFDArecommendations,and correspondingchallengesand goodpractice. Includingdata collection,organ responseandoverall responsederivation,progressionconfirmation,response-basedendpoints,timetoeventendpointslike eventfreesurvivalandstatisticalconsiderations regardinganalysisvisittimewindowandprogramming cutoff | time to response, event-free survival | source | 13 |
| 30 | Clarity: An AI-Powered Shiny Application for Natural Language-Driven Clinical Data Analysis | Jinzhi Zhou, Yusi Liu, Dan Su, Jie Liu |
Exploratory data analysis in clinical trials often demands both deep statistical expertise and strong programming skills, creating a significant bottleneck for cross- functional teams. Clarity (Clinical Analysis & Reporting Interactive Tool with You) is an AI-powered R Shiny application designed to eliminate this barrier by enabling pharmacometricians, biostatisticians, and clinical programmers to perform complex analyses through plain-language instructions - no code required. Clarity integrates large language models (LLMs), including GPT and DeepSeek, through an enterprise LLM gateway to translate natural language requests into executable R code. Users upload clinical datasets (SAS7BDAT, CSV, XLSX) and describe their analytical goals in English or Chinese; Clarity generates, displays, and executes R code in real time, returning publication-ready plots, tables, and summaries within a unified chat interface. Key capabilities include domain-specific skill injection for survival analysis (Kaplan- Meier curves), propensity score matching (PSM/IPTW), efficacy and safety summaries, forest plots, and data cleaning. The selected LLM and token usage are displayed transparently with each response. All generated code is fully visible, editable, and downloadable, supporting reproducibility and auditability. An automated reporting pipeline exports results as HTML or PowerPoint presentations. Clarity demonstrates that conversational AI can serve as a practical coding assistant in pharmaceutical data science, dramatically accelerating exploratory analysis while preserving complete transparency and reproducibility of the underlying R code. Keywords: large language models, R Shiny, clinical data analysis, natural language interface, skill library |
forest plot, clinical data, clinical programming, label statement, adam dataset, proc contents, cdisc | source | 14 |
| 31 | ClaudeAgentSDK and shinyAssistantUI: Extending Claude Code's AI Agent Capabilities to the R Ecosystem | Kaiping Yang, Jie Liu |
Background:ClaudeCode,Anthropic'sAIcodingagent,goes beyondtraditional largelanguagemodel (LLM)APIs bymaintainingpersistentsessions,executingtools,managing files,andcompletingmulti-step tasksautonomously withbuilt-in permissioncontrolsandhook-basedextensibility. Althoughanofficial PythonSDK (claude-agent-sdk-python) exposesthesecapabilities programmatically, noRequivalent existed-asignificantgapforpharmaceuticalRandShiny developers. Methods:Wedevelopedtwo open-sourceRpackages.ClaudeAgentSDK portsthePython SDK, spawningClaudeCode asamanagedsubprocesscommunicatingviaabidirectionalnewline-delimited JSONstreamprotocol.Itsupportsmulti-turnconversations,streamingoutput,sessionmanagement, interrupthandling,toolpermission approval,andhook-basedcustomization,achieving over97% API parityacross699automated tests.Acoro-basedasync architectureensures Shinyintegrationwithout blockingtheReventloop.shinyAssistantUIprovidesthefrontendasanhtmlwidgetwrapping @assistant- ui/react,deliveringstreamingtokendisplay, toolcall visualization,slashcommandmenus,andfile attachmentsviaalightweightShiny module.Backend-agnostic,itconnects toClaudeAgentSDKin afew linesofcode.BothpackagesweredevelopedusinganAI-agent-assistedprogrammingworkflowwith third-partyLLMbackends. Results:ThecombinedstackenablescompleteAIagentapplicationsinRand Shiny.Pharmaceutical use casesincludeclinicaldataqueryassistants,SAS-to-Rtranslationtools,regulatorydraftingaids,and human-in-the-loopanalysispipelines.BothpackagesareopensourceonGitHub. Conclusions:ClaudeAgentSDKand shinyAssistantUIbring ClaudeCode'sfullagentcapabilitiestoR, givingpharmadevelopersaproduction-readystackforAI-poweredShiny applications.Attendees will leavewithworking examplecodeforimmediateadaptation | data cleaning, data science | source | 12 |
| 32 | CodeGenius 2026: Connecting Enterprise Metadata, Standards, Interactive Execution, and AI Coding Agents for Statistical Programming | Jing Wang, Yaohui Zhu, Xuejie Zhou, Jinling Li, You Li |
Asclinicaltrial programmingbecomesincreasingly drivenbystudymetadata,company standards,reusabletemplates,and rapidvalidationfeedback,statisticalprogrammersneed anintegratedworkspacethatextendsbeyond traditionalcodeediting.CodeGeniushas evolvedfromanAI-poweredSASeditor intoanenterprisestatistical programmingworkspace thatbringstogethermetadata,standards,codedevelopment,execution,and AIassistance withinasingleanalysis-scopedenvironment.BuiltonMonacoEditor and integratedwiththe ISPP(IntegratedSASProgrammingPlatform)analysisworkspace,CodeGenius uses backendAPIsandworkspace-level filemanagementtoconnectSAS programs,logs,listings, datasets,studytrackers,standardmacros,templates,andSDTM/ADaM/TLFspecificationsin aunifiedworkflow.Company metadataandprogrammingstandardsarenolongerstatic referencematerials;instead, theydirectlyenablemacrodocumentation,template recommendation,standardheader generation,andcontext-awareprogrammingguidance. CodeGeniusalsosupportsinteractiveSASexecutiondirectlywithintheeditor.Userscan submitthecurrentprogramorselectedcodetoaliveSASsession,monitorexecutionstatus inrealtime,andreviewlogs,listings,generateddatasets,andoutputfileswithoutleavingthe workspace,creatingaclosedfeedback loop fromcodingtoexecutionandreview.Akey advancementinCodeGenius2026is itsAICodingAgentforstatisticalprogramming.Unlikea general-purposechatbot,theCoding Agentoperateswithfullawarenessofthecurrenteditor context,selectedcode,analysis workspace,companystandards,anddomainspecifications. ItsupportsSAScodegeneration,explanation,debugging,loganalysis,andprogram modification.Bycombiningenterprisemetadata,interactiveexecution,and adomain-specific CodingAgent,CodeGenius2026offersapracticaland compliantpath towardagentic statisticalprogramminginthepharmaceuticalindustry | clinical programming, sas enterprise guide, cdisc | source | 8 |
| 33 | Coding Agent Builds BIP Copilot: An LLM-Powered R Shiny Platform for Clinical Biomarker Analysis | Kaiping Yang, Ruiqi Huang, Bohua Chen |
Background:Clinical biomarkerdatainoncology notonlyprovidesinsights ondiseaseriskand treatmentheterogeneity,butaccumulatesacrossstudiesascompany dataassetsformechanism-of- action(MOA) researchandfuturestudydesign.Integratingthesestandardizedresults requireseffective, flexiblesearch.WeintroduceBIPCopilot-anAIassistanttosearchbiomarker-definedsubgroup analysisresultsfromthecompany'stranslational repository, restrictedtovalidatedmethods,unlike generalwebsearchtools. Methods:BIP (BiomarkerInteractiveProfiler) Copilotis anRShinyapplicationbuiltwith theellmer package,connectedtoGPT-4.1viabslib+shinychat.Usersinteractthroughguidedprompts toretrieve resultsor runnewanalyses.QueriesroutethroughaLangChain-poweredRAG layer toregisteredtools coveringSQLexecution,statisticalanalyses,KMsurvival plots,forestplots,boxplots,andbaseline characteristictables.Key featuresinclude BM25/re-rankingfortool-callingaccuracy, dynamiccross- filtering,andon-demandtableformatting.Areporttool serializesconversationhistoryandtool resultsinto structuredRMarkdowndocuments(HTML/PDF/Word).Acodingagentservesas developmentpartner, executingfeaturesunderdeveloper-maintainedinstructions;project-specific skillscodifyreusable patternsfortooldeployment,configuration updates,and taskmigration. Results:Usersposenaturallanguagequestions-from'IsPDL1prognostic instudys1?'to'Whatisthe riskdifference/ratio forORRbetweenPDL1-highandlow?'-andreceivetabular resultswith AIclinical interpretationwithin30seconds.Thecodingagentparadigm reducedfeaturedelivery fromweeksto hourswith noregressiondefects. Conclusions:BIPCopilotdemonstratestwocomplementary advances:apracticalRarchitecturefor LLM-augmentedbiomarkeranalysispreserving statisticalrigorthroughvalidatedtoolfunctions;anda reproducibleAI-assisteddevelopmentworkflowinwhichacodingagentautonomouslyimplements productfeatures-AIbuilding theAIproduct | proc sql, biomarker | source | 15 |
| 34 | Data Collection, Structuring, Analysis, and Reporting Considerations for CRS, ICANS, and IEC-HS in CAR-T Clinical Trials | Zhemin Pan, Jimmy Yang, Shasha Hu, Bo Li |
Cytokine Release Syndrome (CRS), Immune Effector Cell-Associated Neurotoxicity Syndrome (ICANS), and Immune Effector Cell-Associated Hemophagocytic Lymphohistiocytosis-Like Syndrome (IEC-HS) are key toxicities in CAR-T clinical trials and are critical to objective evaluation of product safety and risk- benefit. However, because CAR-T therapies are relatively novel and development experience remains limited, approaches to data collection, data structuring, analysis, and reporting for these events are not yet well standardized. This variability compromises the reliable assessment of these toxicities and limits the interpretability and comparability of safety findings across studies. Drawing on practical project experience and broad literature review, this paper focuses on three key aspects of handling CRS, ICANS, and IEC-HS data: data collection, data mapping, and data analysis and reporting. Relevant information is often distributed across multiple sources, such as adverse event records, dedicated assessment forms, laboratory data, vital signs, neurologic assessments, concomitant medications, and intervention records. A clear and traceable framework is therefore needed to organize clinically related information and support consistent downstream analysis. This paper proposes an end-to-end practical paradigm from data collection to data structure to analysis and reporting for CRS, ICANS, and IEC-HS in CAR-T studies. Key considerations include derivation and presentation of incidence, time to onset, maximum grade, duration, time to resolution, management, and outcome. This framework aims to improve standardization, traceability, and interpretability, thereby supporting more objective safety evaluation of CAR-T products | adverse events, ctcae grade, clinical trial, concomitant medications, safety set | source | 24 |
| 35 | Demotion Trace-Lite: An application for demotion management | Sizhen Chen | In standardized clinical programming environments, system-level code libraries are typically developed within each organization to maximize reuse, consistency, and efficiency across studies. However, these standardized utilities cannot always fully accommodate study-specific requirements through parameterization alone. In such situations, programmers may need to copy selected "system" utilities into the study-level environment and apply controlled modifications to meet specific analysis needs. This paper uses the term "demotion" to describe the deliberate and well-documented customization of standard, system-level utilities at the study level. Importantly, demotions should not be interpreted as coding errors or deviations from best practice, but represent justified, transparent adaptations required to address unique study requirements, while maintaining traceability. From a quality and governance perspective, demotions should be systematically documented to capture how it differs from the original version. Beyond traceability, a well-maintained demotion tracker also enables continuous improvement by highlighting gaps in standard utilities as they are identified in practice. In many cases, however, demotion details remain embedded in code comments and are managed manually, making them difficult to extract, review, or share - especially across teams with different conventions. To address this, we developed Demotions Trace-Lite, a lightweight Python Streamlit-based application that automatically extracts and consolidates demotion information from code. The tool generates standardized, review-ready outputs and captures key metadata such as change rationale and authorship, supporting both ongoing quality improvement and audit readiness | macro definition | source | 7 |
| 36 | Design and Analysis of a Multi-Frequency PRO Collection Strategy in a Clinical Trial | Jiaqi Lu, Zijing Guo |
Thispaper presents apracticalframework forimplementingamulti-frequencypatient-reported outcome(PRO) collectiondesigninclinicaltrials.IllustratedwiththeEORTCQLQ-C30(v3.0)andits subsetIL163,itshowshowfullquestionnairesandsubsets canbeadministered atdifferent frequenciestooptimize clinicaloutcomesand reducepatientburden.Theworkflowcoversscheduling, dataqualitychecks,andCDISC-compliantSDTMandADaMstandardizationandanalysis, underscoringcriticalroles ofthe sponsor,eCOAprovider,andanalyticteam.PracticaldataQC examples andstandardizeddatasetsdemonstratehowmulti-frequencyPRO data canbevalidated andanalyzed withinCDISCstandards tosupportpatient-centered trialdesigns | proc freq, data audit | source | 7 |
| 37 | Designing an ADaM Dataset for Streamlined Drug-Induced Liver Injury Screening Analyses | Ben Gao, Danfeng Fu, Guowei Wu |
The FDA's release of the Standard Safety Tables and Figures - Integrated Guideline (ST&F) in August 2022 introduced default screening analyses for Drug-Induced Liver Injury (DILI). These analyses include the hepatocellular DILI screening plot, the cholestatic DILI screening plot, and a comparison of patients with maximal treatment-emergent liver test abnormalities. Implementing these analyses consistently across studies requires an analysis-ready, standards-based data structure. This paper presents a fit-for-purpose, CDISC-compliant Analysis Data Model (ADaM) Basic Data Structure (BDS)-based dataset design for DILI screening analyses within the FDA ST&F workflow, with an emphasis on programming efficiency, analytical flexibility, and traceability. The proposed ADDILI structure is designed to support the FDA ST&F DILI screening analyses while simplifying derivations and downstream table and figure programming. Concrete data examples illustrate how the proposed structure simplifies the derivation of required endpoints and graphics, reduces repetitive programming, and enhances traceability. Adoption of this ADaM design will promote consistency, interoperability, and regulatory readiness in DILI screening and reporting across studies | adam dataset, trial arms, cdisc | source | 16 |
| 38 | Development of an Efficient Automated Tool for Extracting Variables from Annotated Case Report Forms (aCRFs) and Mapping to SDTM Specification Files | Ziyu Li | Intheprocessofclinicaldatastandardization,creatingandmaintainingStudy DataTabulationModel (SDTM)specificationfilesis acrucial yetoftenmanual task,particularlyrequiringtheaccurateextraction ofvariables andtheir pageinformationfromannotatedCaseReportForms(aCRFs). Thispaper introducesanopen-source,automatedtooldevelopedinPythontoaddressthispainpoint. ThetoolparsesaCRF PDF filescontaininginteractiveannotations,identifies standardizedandcustom variables,andautomaticallyupdates thecorrespondingExcel-basedSDTMspecificationfiles.This includessupplementingmissingvariablesanddomains,andaddingpreciseCRFpagereferences. Throughbatchprocessing andintelligentmatching,thetool significantlyreducesmanual effort-tasks thatpreviouslyrequiredhoursofmanual verificationand entrycannowbecompletedinapproximately5 minutesforpagemappingandvariable/domaincompletion-while improvingtheaccuracyand consistencyofspecificationfilecreation. Thisarticle detailsthetool'sdesignarchitecture,corealgorithms,applicationexamples,andtheefficiency gainsachieved | excel, clinical trial, cdisc | source | 8 |
| 39 | DQO: A Rule Engineering and AI-Driven Platform for Assessing Source Data Quality from SDTM Perspective | Nana Yang, Huan Song, Zhouhao Pan, Chunpeng CZ Zhao |
DQO (Data Quality Overview) is a rule engineering platform implemented in Python with a Streamlit front end, designed to assess source data quality from SDTM perspective. It addresses persistent challenges in clinical data quality workflows-fragmented rule sources, inconsistent implementation, high maintenance cost, and limited reuse-by connecting the full rule lifecycle within a single operational framework: specification intake, Programming Algorithm generation, Python code generation, script execution, standardized output, and diff-based traceability. The framework supports two complementary execution paths. When a validated implementation already exists in the rule library, the framework routes the check directly to that script, preserving consistency and continuity. When no suitable implementation exists, an AI-assisted authoring path generates programming algorithm and candidate Python code from structured inputs-rule description, variable mappings, and dummy example data-before execution and result output. As for this writing, the DQO rule library covers approximately 80 checks spanning most SDTM domains and continues to grow. Large language models serve as controlled drafting accelerators; all generated artifacts remain subject to human review and governance. Unlike tools focused primarily on script execution, DQO emphasizes rule engineering. Programming algorithm function as an interpretable intermediate layer between business rule definitions and executable logic, reducing information loss during translation and strengthening long-term maintainability, reuse, and auditability | database lock, data cleaning, cdisc | source | 11 |
| 40 | DualBookmarker: A Python Tool to Create Dual Bookmarks for Annotated CRF | Xianhua Zeng | The annotated Case Report Form (aCRF) is a required document for FDA submissions that maps each CRF field to its corresponding variable in the SDTM dataset. According to the CDISC SDTM Metadata Submission Guidelines version 2.0 (SDTM-MSG v2.0), the aCRF should be dual bookmarked: (1) bookmarks by chronology and (2) bookmarks by CRF topics or forms, so that reviewers can navigate the document either by visit or by topic. Manually creating dual bookmarks is tedious, time -consuming, and error-prone, particularly for studies with many visits and forms. Existing automation approaches typically rely on commercial Adobe Acrobat plug-ins combined with SAS, spanning multiple environments and requiring paid licenses. This paper introduces DualBookmarker, a pure-Python tool that automatically generates SDTM-MSG v2.0-compliant dual bookmarks for an aCRF using only the open-source PyMuPDF library. It runs on any platform, supports both GUI and CLI modes, and includes built -in QC checks that verify visit count, domain count, and cross-reference consistency before writing the result back to the aCRF. The application is available for download on my GitHub: https://github.com/XianhuaZeng/Projects/raw/master/DualBookmarker.exe | crf, annotated crf, cdisc | source | 4 |
| 41 | DupChecker: A Python Tool for Detecting Code Duplication in Double Programming in Clinical Trials | Xianhua Zeng | In the clinical trial industry, double programming is a key methodology to ensure data accuracy and reliability. However, code duplication between main and QC programming teams undermines the independence required in dual programming. DupChecker is a Python-based desktop application that identifies and reports code duplication between main and QC SAS programs. This paper introduces the tool's functionality, its implementation, and its application in the clinical trial industry. The application is available for download on my GitHub: https://github.com/XianhuaZeng/Projects/raw/master/DupChecker.exe | clinical trial, independent qc, duplicate detection | source | 4 |
| 42 | Efficient Strategies for Handling Data Issues in Clinical Trial Submissions | Kate Sun | Clinical trial data is often complex and prone to issues such as missing data, outliers, and inconsistencies. Efficient identification and resolution of these issues are critical to maintaining data integrity and ensuring successful regulatory submissions. This paper presents a professional SAS framework to proactively identify and address data anomalies in the SDTM.LB domain. By utilizing data visualizations, macros, and practical SAS techniques, this approach offers a systematic solution to enhance data quality and streamline the clinical trial submission process. The paper includes SAS code examples, output visualizations, and guidance for statistical programmers to practice | clinical trial, imputation, outlier, clinical data, missing value, mean imputation, data science, sample, proc sql, proc freq, cdisc | source | 7 |
| 43 | ESTICID: a SAS macro for Estimating Clinically Important Difference using an anchor-based approach in longitudinal data | Zongshi Qin, Yongpei Yu |
Clinicaltrialreports increasinglyrequireaninterpretable linkbetween anumericalchangeina patient-reportedoutcomemeasure(PROM) andthe changethat patients perceiveas important.However, anchor-basedestimationis oftenimplementedinconsistently, particularlywhenrepeated measurements areincomplete oranchorshavemorethan twoordered levels.%ESTICIDisaSASmacrothat operationalizesananchor-basedmixed-effectsmodellingworkflowforlongitudinalPROMdata.The macroacceptsasinglelong-formatdataset,identifies baselinevalues,derives changescores,evaluates theassociationbetweentheanchorandPROMchange,andfitsalinearmixed-effectsmodelusing PROCMIXED.Itsupportscontinuousand categoricalanchors,user-defined referencecategories, directionofimprovement,and covariancestructures.Outputsincludeadjusted estimateswith95% confidenceintervals andforestplotsthat distinguishclinicallyimportantdifferenceswithingroupsfrom differencesbetween anchor-definedgroups.Anillustrativemulticenterrandomizedtrial datasetisusedto demonstratetheworkflow.%ESTICIDisintendedtomakeclinically importantdifferenceestimationmore reproducibleandeasier toimplementinSAS-basedclinicalresearch. Keywords:clinicallyimportantdifference;patient-reported outcome;longitudinaldata;anchor-based | covariance, change from baseline | source | 15 |
| 44 | Evaluating Missing Data Imputation Methods for Clinical Trial Analysis | Jiajun Wen | Missing data is a common challenge in clinical trial analysis and may affect the reliability of efficacy and safety conclusions. Established approaches such as K-nearest neighbors, multiple imputation by chained equations, and other multiple imputation methods are widely used, but their performance can vary when clinical data include nonlinear relationships, high-dimensional covariates, or complex missingness mechanisms. This paper presents a practical comparison of missing data imputation methods under simulated MCAR, MAR, and MNAR settings using public clinical datasets. The evaluated methods include KNN, MICE, standard Generative Adversarial Imputation Network, warm-start GAIN using pre-filled values, and a pseudo-label conditional GAIN approach that incorporates latent subgroup information through clustering. Root mean square error is used as the primary evaluation metric, supported by visual checks of post-imputation data distributions. The results indicate that warm-start pre-filling can improve the stability and accuracy of standard GAIN in simpler missingness scenarios. In more complex MAR and MNAR settings, the pseudo-label conditional GAIN approach shows improved imputation accuracy compared with traditional multiple imputation and standard GAIN across the tested datasets. This work provides a practical evaluation framework for statisticians, clinical programmers, and data scientists who need to compare established and emerging imputation methods for clinical trial analysis. Keywords: Missing Data, Clinical Trials, Multiple Imputation, MAR, MNAR, GAIN, PC-GAIN, Machine Learning | imputation, clinical trial, mean imputation, multiple imputation, data cleaning | source | 12 |
| 45 | Exploring the Architecture of Statistical Programming Languages in the AI Era | Jiang Chao, Wang Li, Wang Juan, Su Ya, Yan Na |
The rapid advancement of artificial intelligence (AI) has fundamentally reshaped the landscape of statistical programming within the pharmaceutical and clinical research industries. The AI-enabled era has catalyzed a shift toward intelligent, automated, and highly diversified workflows. Traditional statistical programming-once dominated by manual coding, repetitive tasks, and rigid SAS-centric toolchains- now faces both unprecedented opportunities and new compliance challenges. In this context, the Shanghai Pharma statistical programming team has explored a modernized programming language framework. By integrating automation, multi-language workflows (SAS, Python, R, VBA), and AI-assisted tools, we have enhanced operational efficiency, output quality, and regulatory robustness. This paper also discusses the evolution of the "SAS Programmer" into a multi-disciplinary Clinical Data Innovator | outlier, csr | source | 8 |
| 46 | From AI-Ready Metadata to Intelligent Agents: Enabling Human-in-the-Loop Coordination in Clinical Programming | Yanan Sui, Tingting Zeng, Yaohui Zhu, Nana Yang, Mingliang Fan |
As AI become increasingly integrated into clinical programming workflows, the effectiveness of automation depends not only on code generation, but also on metadata quality, output transparency, and reliable human oversight. In ADaM development, derivation rules captured in specifications and associated programming comments are often maintained manually and written with varying levels of consistency, which limits their interpretability, reusability, and adaptability in AI-augmented environments. In addition, review and debugging remain time-consuming and cognitively demanding processes, often creating a bottleneck in AI-assisted programming workflows. This paper proposes an integrated framework that links AI-ready metadata with human-in-the-loop automation for intelligent ADaM program development, review, and debugging. Using real-world ADaM examples from ACIRA (ADaM Code Intelligence for Rapid Analysis), we demonstrate how structured metadata enables derivation interpretation, code generation, and lifecycle adaptation as study requirements evolve. Building on this foundation, the framework introduces review and debug agents to support quality assessment and issue resolution: the review agent provides structured feedback and validation outcomes to support logic verification and result evaluation, while the debug agent analyzes execution logs and generates context-aware correction suggestions. Together, the proposed metadata-centered framework with agent-assisted mechanisms provides a scalable approach to intelligent clinical programming automation. It improves the consistency and maintainability of ADaM derivation logic, enhances reliability of automated outputs, and ensures human supervision at critical decision points. More importantly, it demonstrates that sustainable AI integration in clinical programming requires both machine- readable metadata and interactive mechanisms for continuous review, debugging, and refinement | macro parameter, response variable | source | 6 |
| 47 | From SAP to TFL: An AI-Assisted Workflow Framework for Automated Shells and TFLs | Xiaoyu Ding, Guoqing Zhao, Huafeng Gao, Zhiyuan Guo |
The core deliverables in clinical trial statistics and programming typically include the Statistical Analysis Plan (SAP), statistical analysis shell templates (Shells), analysis programs, standardized macro libraries, controlled terminology, SDTM and ADaM datasets, and the final Tables, Figures, and Listings (TFLs) for clinical study reports and regulatory submissions. In the traditional workflow, programmers must repeatedly read SAPs, Shells, data specifications, and project communications to identify requirements, then manually translate these into table of contents (TOC) entries, program parameters, output formats, and validation checklists. While controllable, this process is costly, and frequently leads to missed requirements, version inconsistencies, duplicated development, and tracking difficulties-especially when managing multiple projects, frequent SAP amendments, or multi-team collaboration. This paper presents an AI-assisted workflow framework designed for clinical statistical programming. The framework uses a structured metadata approach in which a large language model (LLM) extracts three categories of key metadata from the finalized SAP: table content metadata, table structure metadata, and table programming metadata. Through rule validation and human confirmation, these are consolidated into versioned project metadata. Confirmed metadata then drives the Tracker, standardized Shell Library, controlled terminology, macro library, and template mapping rules to enable semi-automated generation of Shells and TFL programming artifacts. This paper discusses the framework's process design, key metadata model, the interplay between Tracker and Library, quality control, validation strategy, and implementation risks | data cleaning, adverse events, analysis set, clinical programming, safety population, data science, controlled terminology, adverse event summary, system organ class, treatment-emergent adverse event, missing value, efficacy summary, class variable, in=, quality control, adam dataset, clinical trial, cdisc | source | 18 |
| 48 | From Spreadsheets to Smart Platforms: A Framework for Modernizing SDTM Specification Management Using the Microsoft 365 Ecosystem | Shuting Lou | Inclinical programming,MicrosoftExcelhaslongbeentheprimarytool formanaging studydeliverables suchasSDTMmappingspecifications,trackinglogs,andoperationaldocuments.Despiteits flexibility, spreadsheet-basedworkflowspresent persistentlimitations suchasfragmenteddataacrossmultiple files, limitedsearchabilityandmanual versioncontrol. Thispaper presents aframeworkfortransitioning fromspreadsheet-basedprocessestoanintegrated approachusing theMicrosoft365ecosystem -SharePointLists,PowerApps,andPowerAutomate- asano-code/low-codealternativeanddemonstrateshowthesecomponents combinetocreatea structured,interactiveworkflowenvironmentforclinicalprogrammingteams. Inthispaper,we presentaSharePoint-basedSDTMSpecificationExplorer.SharePointLists replace EXCEL-basedmetadatastoragewith structured,filterable,version-controlled specificationdata.Power Appsprovidesanintuitiveinterfacethatenablesusers tobrowseandfilter acrossdomains,variables,and SDTMIGversions,whilealsosurfacingcross-domainrelationshipsand mappingdependencies thatare difficulttotraceintraditional multi-tabworkbooks.PowerAutomateintroduces event-drivenautomation suchasnotificationstriggeredby specificationupdates,keepingall teammembersalignedin realtime. Withthisplatform,itisalsopossibletoextractadefine-ready specificationin one-click. Thispaper highlightshowleveragingexistingenterprisetoolscansignificantlyimproveefficiency, traceability, andusabilityinclinicalprogrammingworkflows,withoutrequiring customsoftware development.Theproposedapproachprovides ageneralizablemodelformodernizingspreadsheet- drivenprocessesacrossdifferentstudiesandorganizations |
data cleaning, controlled terminology, cdisc | source | 14 |
| 49 | Handling Missing Data in Analgesia Clinical Trials and Corresponding SAS Implementation | Jingbo Zhu, Jiehan Zhang |
In analgesic clinical trials, primary endpoints are mostly patient-reported pain scores, such as Numeric Rating Scale (NRS), Visual Analogue Scale (VAS), Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) etc. Missing data is prevalent, mainly caused by premature discontinuation, use of rescue therapy, adverse events, poor participants' compliance, or missed assessments. Such incompleteness increases the complexity of statistical analysis and may lead to biased estimates of clinical efficacy. This paper addresses missing data handling in analgesia clinical trials with repeated measurements and longitudinal observations. We focus on two primary methods under the Missing at Random (MAR) assumption-Multiple Imputation (MI) and Mixed-Effect Model Repeated Measures (MMRM), also review Windowed Last Observation Carried Forward (wLOCF) as a practical imputation approach for rescue therapy. By bridging statistical methodology with practical SAS implementation, this paper helps clinical programmer to handle missing data more effectively in analgesia clinical trials | imputation, missing value, covariance, proc step, fixed effects, proc mi, multiple imputation, proc mianalyze, proc mixed, clinical trial, repeated measures, covariance structure, sensitivity analysis, likelihood function, change from baseline, least squares mean, missing not at random, ancova, goodness of fit, inferential statistics, sample, adverse events, maximum likelihood estimation, estimate, data science, treatment effect, longitudinal data, correlation, regression analysis, restricted maximum likelihood, confidence interval, unstructured covariance, last observation carried forward | source | 17 |
| 50 | Health Technology Assessment (HTA) Under EMA Framework: From Emerging to Mandatory Submission | Jun Zhai | Health Technology Assessment (HTA) is a multidisciplinary process that evaluates the clinical, economic, and social implications of health technologies to inform coverage and reimbursement decisions. The European Medicines Agency (EMA) has undergone significant shifts in its HTA alignment strategy, particularly following the implementation of the EU HTA Regulation (EUR) 2021/2282. Under the regulation, the HTA coordination group and its sub-groups, composed of national HTA bodies jointly conduct Joint Clinical Assessments of new medicines and certain high-risk medical devices, starting with new cancer medicine and advanced therapy medicinal product. This paper examines current landscape of HTA submissions under EMA framework, identifies gaps and argues clearer direction for statistical programmers' involvement and responsibilities in HTA submission | subgroup analysis, clinical trial, treatment effect | source | 7 |
| 51 | Implementation of IPCW in Oncology Trials and Enhanced SAS Macro Tools | Jiayi Chen, Meiqing Huang |
Treatmentcrossover fromthecontrolarmtotheexperimental armiscommoninoncology clinicaltrials andcancomplicateinterpretationofoverall survival(OS).Inverseprobabilityofcensoringweighting (IPCW)isarecognizedapproachforestimatingahypotheticalno-switchingtreatmenteffect,but itsusein productionclinicalprogrammingrequires morethanastatisticalmodelstatement.Programmersmust prepareinterval-leveldata,aligntime-dependentcovariates,estimateandassessweights,fitweighted survivalmodels,andgenerate uncertaintyestimatesthatreflectthefullanalysisworkflow. Thispaper presents anenhanced SAS macroframeworkforIPCWimplementationin oncologytrials. AlthoughseveralRpackagesprovideusefulmethodologicalreferences,SAS-nativeimplementation remainsimportantinclinicaltrialreportingenvironments.Buildingonthe SAS macroframeworkdescribed byMosier(2023),weproposeenhancementsfortime-dependent covariatehandling,quality control,and bootstrapconfidenceintervalestimationfortheadjustedhazardratio.Theresultingworkflowis intended tomakeIPCW analysesmoretransparent,reproducible,andpractical forstatisticalprogrammingteams. Wealsodevelopedanindependent RscriptofthecompleteIPCW workflowtoassessconsistency betweentheRandSASimplementations | bootstrap, quality control, confidence interval, proc calis, proc phreg, sample | source | 13 |
| 52 | Implementing a Single-Agent Bayesian Logistic Regression Model in SAS for Dose Escalation in Phase I Oncology Trials | Zhanwen Peng | TheBayesianlogisticregressionmodel (BLRM)isawidelyusedmodel-baseddesignfordoseescalation inphaseIoncology trials;nevertheless,reusableSASimplementationsthatmakeMCMCchecking explicitareseldomreported inconferencepapers.This paperdescribes apragmaticSASworkflowfora single-agentBLRM.TheBLRMmodels dose-limitingtoxicity (DLT)counts usingabinomial likelihoodand alog-doselogisticregression,assigns abivariate normalpriortotransformedmodelparameters,and appliesescalationwith overdosecontrol(EWOC)tosupportdose recommendation.Theimplementation isstructuredasmodularSASmacrosforsingle-chainposteriorsampling,multi-chainformalposterior inference,posteriortoxicitysummarization,EWOC-based decisionsupport,single-trialexecution, repeatedsimulation,andsimulation summarization.Inaformalmulti-chainanalysis,thewrapperruns independentlyinitializedPROCMCMCchains,retains within-chaineffectivesamplesize(ESS),Monte Carlostandarderror(MCSE),autocorrelation,Geweke,andtrace-review outputsforeverychain,and calculatestheclassicalGelman-Rubin R-hatasabetween-chaindiagnostic.Pooledposteriordraws are usedforformaltoxicity summariesand EWOCdecisionsupportonly afterboth levelsofdiagnosticreview aresatisfactory. This separationretainsacomputationallyefficientsingle-chainengineforsimulation whileprovidingatransparent,auditable,andreusableBLRMworkflowforformalinterimanalysis. Keywords:BLRM,doseescalation,EWOC,PROCMCMC,MCMCdiagnostics,R-hat,simulation | dose escalation, logistic regression | source | 11 |
| 53 | Interactive Clinical Plot Gallery with Cross-Plot Subject Linking in R Shiny | You (Lea) Lyu | Clinical trial data visualization plays a critical role in understanding treatment effects, safety profiles, and individual subject responses. Traditional static reporting tools often lack the interactivity needed for exploratory data analysis during study review. In this paper, we present a user-driven clinical plot gallery built with R Shiny, featuring eleven plot types commonly used in oncology clinical trials: line plots, waterfall plots, scatter plots, box plots, spider plots, swimmer plots, multi-Y plots, tornado plots, bar plots, Kaplan-Meier plots, and forest plots. A key feature of this plot gallery is the cross -plot subject linking system, which allows users to highlight individual subjects across all visualizations simultaneously, enabling seamless exploration of subject-level data. The plot gallery also supports a user-configurable layout where users can select and arrange plots of interest side by side. A pattern -based highlighting system enables users to identify and visually emphasize subgroups of subjects sharing clinically meaningful response patterns across compatible plots. A click-to-explore feature on selected plots triggers a subject profile modal displaying laboratory, adverse event, and best overall response data. A subject journey timeline provides a longitudinal, event-level view of each highlighted subject's clinical course. An integrated statistical options panel allows users to overlay hypothesis test results, effect sizes, and multiple comparison adjustments directly on selected plots. Additionally, a per-plot code editor powered by the shinyAce package allows users to modify, preview, and apply custom plot functions directly within the dashboard. This paper describes the design, implementation, and key technical considerations of building such a user-driven platform using R Shiny and the shinydashboard, plotly, shinyWidgets, and shinyAce packages | adverse events, clinical trial, best response, scatter plot, waterfall chart, change from baseline, clinical data, data visualization, effect size, box plot, swimmer plot, ctcae grade | source | 10 |
| 54 | Interactive Clinical Trial Data Review Using Non-Shiny R HTML Dashboards | Peng Tan | Clinicalprogrammingteamsarefrequentlyaskedtosupportdata reviewandriskmonitoringforstudies thatspandozensofcountries andhundreds ofsites,withsourcedata storedinSASdatasetsand review decisionstrackedacrossmultiple functions.Static tablesand listingsanswer specificquestions,butthey rarelyhelpteamsseethebiggerpicture-whichsitesaredrivingmostofthe burden,whether thereview backlogisgrowingor shrinking,or howthis month'ssnapshotcompares tothelastone. Thispaper describes anon-ShinyHTMLdashboardapproachbuiltwithRand JavaScriptforprotocol deviation(PD)datareview.TheapproachusesRMarkdownandflexdashboardtorenderaself- containedHTMLfilefromSAS sourcedata,with allchartrenderinganduser interactionhandledclient- sideusingPlotly.jsand vanillaJavaScriptmodules.Theoutputrequiresnoserver,noShinyruntime,and nospecialviewer-itopens inanymodernbrowser. Beyondthe basicarchitecture,thepaperdiscusseshowthedashboardhandlesmulti-audiencereporting needs:executiveKPIviews,site-levelcross-filtering,country burdencomparisons,editablereview annotations,andspreadsheetupload againstafixedHTMLshell.Italsointroduces abrowser-based local LLMmodule thatcangeneratenatural-languagesummariesfrompre-aggregateddashboardstatistics, implementedusingtheMLCAIWebLLMlibraryandrunningentirely offlinethroughWebGPU. By executinginferencelocallyontheuser'sdevice,thisapproachsimultaneouslyaddressestwopersistent challengesin clinicalAI adoption:data privacy-nopatient-level orstudy-level informationleavesthe localmachine-andinfrastructuredependency -nocloudAPIkeys,noserver provisioning,andno networkconnectivityrequiredatruntime. Thegoalofthispaperisnottoadvocateforanyparticular toolset,buttoshareapracticaldesign approachthatmany Rprogrammingteamscanadaptwithintheirexistingclinical workflowwithout adoptingahostedapplicationplatform | clinical trial, stream processing | source | 7 |
| 55 | Leading Through Complexity: A Compound Lead Perspective on Planning, Coordination, and Quality Control | Fuyuan Pang | In complex clinical trial settings, the Compound Lead role extends well beyond managing programming deliverables. It requires proactive planning, cross-functional coordination, resource balancing, and practical quality leadership. Drawing on lessons from a recent project, this presentation shares a leadership perspective on how statistical programming leads can operate effectively in demanding study environments | quality control, database lock | source | 4 |
| 56 | Making Composite Endpoints Work: An ADaM Programming Approach to Sustained Remission | Qing Xue, Fen Wu, Shuo Zhang, Peixuan Wang |
Compositeandsustainedbinaryendpoints arecommoninPhaseIIItrialsbutaredifficulttoimplement becausethey combinemultipleclinicalconcepts,longitudinalrules,missingdata handling,and estimand requirements.This paperpresentsalayeredADaMprogrammingframeworkforderivingsustained remissionendpointsinanADEFF dataset.Theapproachseparatesderivation intocomponent-level rules, visit-levelcomposite flags,andfinalsustainedendpoints whileincorporatingtime-awarelogicand estimand-alignedhandling ofmissingdata andintercurrentevents.Theframeworkimproves traceability, reviewability,validationefficiency,andreuse,anditcanbeappliedtoothercomplexlongitudinal compositeendpoints. Thispaper presents alayeredADaMprogrammingframeworkforderivingsustainedremissionendpoints inanADEFFdata set.Themethod separates derivationinto component-levelrules,visit-levelcomposite flags,andfinalsustainedendpoints,whileincorporatingtime-awarelogic andestimand-alignedhandling ofmissingdataandintercurrentevents.This designimprovestraceability,reviewability,validation efficiency,andreuse.Althoughillustratedwithsustainedremission,theapproachisapplicabletoother complexlongitudinalcompositeendpoints | intercurrent event, data cleaning, cdisc | source | 15 |
| 57 | Modernizing the Vaccine TAUG: An End-to-End, Output-Driven Framework for Reactogenicity, Immunogenicity, Efficacy, and Reviewer-Ready Submissions | Lin Chen, Jialiang Liu |
The Clinical Data Interchange Standards Consortium (CDISC) Therapeutic Area User Guide for Vaccines (TAUG-Vax) v1.1 (9 April 2018) remains a useful reference for reactogenicity data, but it was issued as a provisional, Study Data Tabulation Model (SDTM)-oriented guide built on the foundational standards of that period (SDTM v1.4 / SDTM Implementation Guide [SDTMIG] v3.2). Contemporary vaccine programs are far broader: multi-dose and booster schedules, multivalent and co-administration designs, adaptive master protocols, humoral and cell-mediated immunogenicity, disease-endpoint efficacy, and immunobridging. CDISC foundational standards have also moved on - SDTMIG v3.4 revised the Immunogenicity Specimen Assessments (IS), Laboratory Test Results (LB), and Microbiology Specimen (MB) domain scopes and added the Cell Phenotype (CP) and Genomics (GF) Findings domains [2] - and U.S. Food and Drug Administration (FDA) Center for Biologics Evaluation and Research (CBER) Office of Vaccines Research and Review (OVRR) technical expectations have become more specific [3]. As a result, v1.1 is no longer sufficient as an end-to-end implementation guide. This paper identifies the main gaps in TAUG-Vax v1.1 across reactogenicity, immunogenicity, and clinical endpoints (efficacy and immunobridging), then proposes a modernized, COVID-19-TAUG-style guide organized as an output-driven framework in which implementation decisions are traced from protocol and Statistical Analysis Plan (SAP) through Clinical Data Acquisition Standards Harmonization (CDASH), SDTM, and Analysis Data Model (ADaM) to Clinical Study Report (CSR) tables, registry and lay outputs, reviewer guides, and validation. A worked example follows one solicited injection-site reaction from source data to a reviewer-traceable tables, listings, and figures (TLF) cell | geomean, analysis set, data cleaning, proc hpsplit, adverse events, adam dataset, reviewer guide, sdtm domain, sdtm programming, clinical trial, cdisc | source | 18 |
| 58 | One Pipeline, Zero First Drafts: A Metadata-and-LLM Toolkit for Automated TFL Program Generation | Hefan Jing, Peilin Liu |
Inclinical trialprogramming,producing Tables,Figures,andListings (TFLs)typicallybegins witha manual,labor-intensivestep:reading aTableofContents (TOC),selecting theappropriatesystemmacro foreachoutput,andfillinginthecorrectmacroparameters.Thispaperpresents generate_pdt,aPython- basedpipelinethatautomates this"first-draft"process.Thecoreidea istodecomposeeachTOCentry intothreestructuredcomponents-analysis type,filterconditions,andinputdataset-thenmatchthese componentstotheappropriatesystemmacrousingathree-levelstrategy: exactmatching,synonym matching,andsemantic matchingviaalarge languagemodel(LLM).Thetoolgenerates acomplete ProgramDeliverables Tracker(PDT)andready-to-runSASprogramswithout anymanualintervention. Wedescribethedecompositionlogic,theknowledgebasestructure,thethreematching methodsand theirrespectivestrengthsandlimitations,anddiscussthetool'scurrentgaps andfuturedirections | safety population, adverse events, analysis set | source | 10 |
| 59 | R in study validation | Liang Han | In recent years, the pharmaceutical industry has increasingly begun exploring the use of R to replicate and perform tasks traditionally handled by SAS, driven by the rapid growth of the open -source ecosystem and the demand for analytical flexibility. While R offers great flexibility, it brings two main challenges: maintaining strict regulatory compliance in clinical studies and overcoming the steep learning curve for SAS programmers. To solve these issues. This presentation introduces a hybrid programming strategy that integrates SAS and R to leverage the strengths of both languages: SAS for core study execution to guarantee data accuracy and regulatory submission readiness, and R for quality control (QC), exploratory analysis, and enhanced workflow flexibility. This hybrid approach balances the regulatory robustness of SAS with the flexibility and analytical power of R, providing a practical, scalable solution for clinical study validation that enhances efficiency while maintaining compliance.Beyond proposing this integration strategy, a core objective of this work is to help SAS programmers overcome the learning curve of R, enabling them to quickly master its application and seamlessly apply R within real -world clinical projects. To facilitate the transition for SAS programmers, a structured training framework is proposed, encompassing knowledge acquisition, team upskilling, real-world project practice, internal knowledge sharing, and targeted guidance for SAS-to-R adoption. We developed an LLM-based application that automatically converts SAS code into R code. This tool allows programmers to instantly see R equivalents of their SAS logic, helping them learn R through practical use in real projects | sas programmer, data cleaning, clinical programming, clinical trial, quality control | source | 7 |
| 60 | R Submissions in Practice: The J&J Implementation and Emerging AI Opportunities | Wenlan Pan | The question for R-based regulatory submissions is no longer whether they work, but how to make them work reliably, repeatedly, and at scale. With successful filings now reported across multiple sponsors and health authorities, the field has the evidence needed to move from feasibility to execution, focusing on implementation decisions, documentation standards, and where generative AI may responsibly reduce effort. This paper addresses those questions in two connected parts. First, it documents the Johnson & Johnson submission experience in depth, covering package management and governance, ADRG design, eCTD file structure adaptations, mock submission practice, and adaptations for multi-regional filings including China's CDE, with key enablers described as reusable design patterns applicable across sponsors and therapeutic areas. Second, it explores where generative AI may fit within the submission programming workflow to reduce manual effort in repetitive and documentation-heavy tasks, and discusses the controls needed to keep AI-assisted work compliant and reviewable, including traceability, reproducibility, and human oversight requirements. By combining an in-depth implementation case study with an exploratory AI assessment, this paper provides a practical reference for teams building scalable R-based workflows and a forward-looking perspective on how AI may enhance efficiency without compromising regulatory compliance. It is designed to be read alongside the companion paper on health authority landscape analysis and AI application frameworks. | regulatory submission, data cleaning | source | 11 |
| 61 | Safety Estimands and Analysis Dataset Design: Advancing Safety Evaluation in Clinical Trials | Kun Ren, Zeyu Guang |
With the adoption of the ICH E9(R1) estimand framework (ICH 2019), there is increasing interest in applying estimand principles to safety evaluation in clinical trials. Traditional safety analyses are often descriptive and may not explicitly define the safety question or treatment effect of interest, particularly in the presence of intercurrent events such as treatment discontinuation, dose modification, rescue medication, or death. This paper introduces the concept of safety estimands, outlines the background for their use, and discusses their application in safety analyses. It also highlights key considerations for analysis dataset design within the CDISC ADaM framework (CDISC 2016), with particular focus on time- to-event dataset structures to support estimand-aligned derivations and transparent handling of intercurrent events. Integration of the safety estimand framework with the aligned ADaM design strengthens the scientific and operational quality of safety analyses | adverse events, intercurrent event, adam dataset, serious adverse event, treatment effect, concomitant medications, hypothetical strategy, while-on-treatment, treatment policy strategy, randomization, clinical trial, proc means, population, time-to-event endpoint, data science, cdisc | source | 11 |
| 62 | ShellPainter & CDARS: An AI-Powered Digital Framework for Integrated shell-TFL Automation | Jiapeng He, Mingliang Fan, Yanan Sui, Lin Jiang, Mingfeng Zhou, Xuejie Zhou |
Background: Tables, Figures, and Listings (TFLs) creation remains a labor -intensive process requiring significant coordination between statistical programmers, statisticians, and medical scientists. Traditional workflows involve manual shell design, repetitive parameter entry, substantial programming effort for each deliverable, and can't guarantee the consistency between shell and final TFLs. Objectives: We present ShellPainter and CDARS (Clinical Data Analysis and Reporting System), an integrated web -based platform that transforms shell & TFL development through intelligent metadata management and AI-assisted automation. Methods: ShellPainter provides an intuitive interface for designing study level master TFL shells, automatically capturing structural metadata (e.g., column hierarchies, table block definitions, and statistical model configuration), and tracking shell ongoing changes. CDARS extends this foundation by linking shell metadata with SDTM/ADaM annotations, enabling automated TFL generation & validation through SAS & R programs. The system leverages large language models (LLMs) to auto-select analysis contents in mock shell based on protocol/SAP, intelligently suggest variable/parameter mappings & analysis filters, auto debug & fix TFL metadata/program. Version control with full audit trails ensures regulatory compliance, while real-time visualization features streamline the review process. Results: Implementation across 60+ clinical studies demonstrated 70% resource reduction in TFL delivery and significant decrease in programming errors. The AI -assisted metadata generation achieved 80 -90% accuracy in suggesting appropriate analysis variables, parameters, and filters. Conclusions: ShellPainter and CDARS represent a paradigm shift from document -centric to metadata - centric TFL development, enabling reproducibility, traceability, and substantial efficiency gains while maintaining the flexibility required for diverse analysis & reporting needs | data cleaning, mock shell, cdisc | source | 7 |
| 63 | Stop Building, Start Prompting: Leveraging LLM-Native Capabilities for TFL Automation with Minimal Scaffolding | Zhongqian Xu | Clinical TFL generation traditionally requires substantial study-specific engineering, including Word-shell parsing, variable mapping, statistical programming, and output formatting. This paper presents ADaM2TFL, an exploratory design and working prototype in which a compact Markdown "skill file" provides an LLM coding agent with project context: directory layout, shell conventions, statistical rules, formatting constraints, and known failure patterns. Reusable deterministic utilities support shell normalization, metadata caching, and structural validation, while the agent interprets these artifacts and generates candidate R programs and RTF outputs. The objective is not to eliminate software, but to minimize per-study custom scaffolding. A five-phase workflow combines configuration, shell confirmation, metadata caching, code generation, and verification, with human review retained at key decision points. The design also includes an experimental knowledge-accumulation loop and an exploratory shell-free mode; both are target capabilities rather than validated production features. Experience with the prototype suggests that many failures can be addressed by improving the governing context instead of adding study-specific code. The central proposition is therefore that, for LLM-native clinical reporting, part of the engineering bottleneck shifts from building automation infrastructure to authoring precise, reviewable context | data cleaning, audit trail | source | 6 |
| 64 | TFL Generation Using R: Negative Binomial Regression Implementation and Comparison with SAS | Ye Sun, Jianchang Hu |
Tables,Figures,andListings (TFLs)arefundamentalforreporting andinterpretingclinicaltrialdata in regulatorysubmissions andaretraditionallyproducedusing SAS.WithincreasinginterestinRasan open-sourcealternativeforclinicaltrial analyses,thispaperprovidesacomprehensivecomparisonofkey featuresbetweenSASandRimplementationsfornegativebinomial regression. Thepaperdemonstratesthat,bymanually adjustingtheestimatedvariance-covariancematrixinR,exact matches(tothe0.001level)forparameter estimates,least-squares means,confidenceintervals,andp- valuescanbeachievedusingR'sglm.nbfunctioncomparedtoSAS'sPROCGENMODprocedure. Specialattentionisrequiredforthedispersionparameter duetodifferingparameterizations betweenthe twolanguages.Additionally, theuseofthegtsummarypackagein Rforgeneratingformattedoutputis illustrated | regression analysis, negative binomial regression, binomial distribution, covariance | source | 11 |
| 65 | The Effective Preparation of NMPA Submission Package | Fu Wang | This paper outlines the procedural steps, roles, and responsibilities integral to preparing the National Medical Products Administration (NMPA) submission package, which is based on the translation of foreign language source materials. The document presents a comprehensive overview of quality assurance and proofreading protocols applied to vendor-translated content. Additionally, a Visual Basic for Applications (VBA) tool has been developed and implemented to enhance the efficiency of quality review processes and facilitate automated corrections for standardized dataset and variable labels where applicable. The paper details the underlying logic and functionality of this VBA tool. Furthermore, critical quality assurance considerations are discussed to ensure the integrity and completeness of the NMPA submission package | excel, label statement, data cleaning | source | 6 |
| 66 | The reporter package: A Powerful Reporting Package to Replicate SAS Output | Bill Huang | With the increasing application of the R programming language in clinical trial research, a primary objective is to successfully replicate SAS output results. The "reporter" package represents a comprehensive and user-friendly reporting solution, offering capabilities comparable to SAS's PROC REPORT/ODS. This package supports multiple output formats including TXT, PDF, RTF, DOCX, and HTML-while efficiently handling large-scale data. Leading pharmaceutical companies, such as Amgen, have adopted "reporter" as a key component of their reporting workflows. This paper provides a systematic introduction to the fundamental usage of "reporter", covering applications from basic to advanced levels. Furthermore, it demonstrates how "reporter" can generate outputs nearly identical to SAS through concise code, illustrated by standard clinical trial tables and listings. Finally, this paper highlights useful features such as break labels and group cohesion, and outlines the future development roadmap, including RTF code insertion |
clinical trial, create table, demographics table, shift table, sas output, system organ class, proc report, sample, proc step | source | 24 |
| 67 | Trace Data Framework for Metadata Network, Impact Analysis, and Visualization Across RAW, SDTM, ADaM, and TFL | Nana Yang, Tingting Zeng, Yaohui Zhu, Yanan Sui, Mingliang Fan |
End-to-end traceability across rawdata, SDTM, ADaM, and TFL is an important foundation for lineage explanation, impact assessment, and review preparation in clinical data development. In practice, however, relevant relationships are often distributed across specifications, mapping logic, derivation logic, reviewing outputs, and reporting artifacts, making cross-layer dependencies difficult to understand as a whole and downstream impact scope difficult to assess in time. To address this problem, this paper presents a trace data framework that organizes distributed cross-layer relationships into a queryable metadata network. The framework generates trace segments from three linkage domains, RAW-to- SDTM, SDTM-to-ADaM, and ADaM-to-TFL, and dynamically assembles them into continuous lineage chains according to user-selected scope and filters. Using this framework, users can identify which CRF fields support a given TFL output, analyze the upstream origins of an ADaM derivation, assess downstream impact caused by SDTM changes, and inspect the relationships between CRF forms and reporting TFL targets. To improve practical usability, the framework further provides Sankey-style flow diagrams, trace tables, form-to-output mappings, and downloadable trace datasets for review, communication, and documentation support. Practical implementation demonstrates that once trace relationships are organized into navigable, composable, and reusable trace data, they transcend their role as a governance requirement. Instead, they become a computational layer that actively supports explainable, auditable, and maintainable clinical data development | clinical data, bds, cdisc | source | 7 |
| 68 | Validation Strategy for TLF AI-Generated Results | Han Wang, Zongfeng Lei, Patick Zhang |
Context. Generative AI is now practical enough to support parts of TLF programming. Our motivation is simple: produce more tables with less repeated manual effort, shorten turnaround time, and still keep quality visible to reviewers. But once AI starts drafting SAS code, the key question becomes how to confirm that the code follows the shell, uses the right study data, and is suitable for statistical programmer review. TLF programming is a good setting for this question because the work already depends on familiar clinical reporting materials-shells, study specs, and runnable SAS macros-that can be checked step by step. Design logic. Evaluation should follow the path of the generated code. First, assess whether the system understood the shell correctly. Next, assess whether the code uses the right ADaM data, variables, filters, and treatment groups. Then assess whether the SAS program is complete, runnable, and understandable for a statistical programmer. These checks inform-not replace- normal programming review and QC. To make assessment repeatable at scale, we use evaluation harnesses: automated infrastructure that applies the same checks on benchmark cases at each release. Evaluation runs in a separate environment from study delivery: changes to generation settings do not affect TLFs already delivered for a study, and production approval stays the same. This paper shares our overall evaluation strategy and framework. It does not describe internal scoring rules, benchmark contents, or implementation details. The aim is practical guidance for teams adopting assisted TLF programming-not a regulatory validation package or a technical specification. Keywords: TLF generation, AI-generated SAS, evaluation, human review, statistical programming | statistical programmer, clinical research | source | 7 |
| 69 | When Data Capture Becomes Data: How ePRO and e-Diary Systems Reshape CDISC Standards, Statistical Programming, and Endpoint Analysis in Vaccine Studies | Jialiang Liu, Lin Chen |
Electronicclinicaloutcomeassessment(eCOA) systems,includingelectronic patient-reportedoutcome (ePRO)e-Diaries,areincreasinglyusedtocollectvaccinesolicitedreactions.Unlikepaperdiaries,they canprovide trustedentrytime,uploador receipttime,collectionmode,devicecontext,andaudit metadata.Theseelements mayarrivethroughdifferentextractsandatdifferentgrains.Theanalysis-and traceability-relevantsubsetshouldthereforeberetained withoutreproducingthefullaudittrail in submissiondatasets.This paperfocusesonfourconsequences:distinguishingatruenegativefroma missingassessment;separating clinical,intended-day, entry, andreceipttimes;preservingsubjectand investigatorrecords;andderivingquestion-specificdenominators.Afive-subjectDays 1-8example separatesexpecteddiaryparticipation (N_DIARY),endpoint-specificevaluability(M),electronic submission,andcumulative timingwindows (≤24,≤48,and≤72hours).Italsoshowshowpaperfallback cancontributeclinical evidencewithout electronictiming.Weproposeanexpected-assessmentstaging frame,selectedSDTMmetadata,explicitADaMflags,and reviewer-facingshellsforparticipation,timing, andmaximumseverity. Theexamples areconstructed;timing-restrictedviewsaresupportiveor sensitivityanalysesunlessotherwisepre-specified. Keywords: eCOA, ePRO, e-Diary, Vaccine Reactogenicity, SDTM, ADaM | reviewer guide, intercurrent event, cdisc | source | 16 |
| 70 | Word2PDF: A Python Tool to Convert and/or Combine RTF Files into a Single Bookmarked PDF/RTF with Table of Contents in a User-defined Order | Xianhua Zeng | As part of clinical trial reporting, a large number of RTF/PDF outputs are created, and at the completion of a milestone in a study, it is required to convert and/or combine all reports into a single, reviewer -friendly document for easy delivery and review. This paper presents a Windows desktop application built with Python to convert and/or combine multiple RTF/PDF files into a single bookmarked PDF/RTF/DOCX document with a table of contents in a user-defined order. The application, called Word2PDF, runs on Microsoft Windows 10 or later with Microsoft Word installed. It also supports DOC and DOCX files as input. Key features include navigation heading insertion and per-section page numbering. The app and test files are available on my GitHub: https://github.com/XianhuaZeng/Projects/tree/master/Word2PDF | clinical trial, excel | source | 7 |
This site serves as a convenient hosting location and does not work as a publisher. All copyrights belong to their respective owners.
The information on this site is provided for general reference purposes only. Whilst every effort is made to ensure that the site is up to date and accurate, no warrant, nor any responsibility or liability for, the accuracy or completeness of the content or for any loss which may arise from reliance on information contained in this site is provided. Links to other sites are provided for your convenience but no responsibility or liability for the content of those sites or of any external site which links to this site is accepted. References to any specific commercial product, process, service, or corporation are also for information only, and do not constitute endorsement, recommendation or favoring.
SAS is a registered trademark of SAS Institute Inc. SAS, Statistical Analysis System, and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA registration.