Poster Sessions

Lightning talks and poster viewing take place during the conference. Posters are displayed in the Stanford room (see the Venue floor map).

BBSW 2026 Poster Awards

Congratulations to our winners!

Thank you to every presenter for sharing your work, and to everyone who stopped by and voted.

Accepted Posters

P01

When Interim Overall Survival Signals Suggest Harm: A Framework for Decision-Making Under Uncertainty

Andrew Mills
Read Abstract ▼
Interim overall survival (OS) analyses in oncology can be difficult to interpret when progression-free survival (PFS) is favourable but early OS data highlight potential harm. Such signals may reflect immature follow-up, informative censoring, post-progression therapies, or true harm, creating uncertainty for sponsors, regulators, and trial teams. Building on the work of a multi-stakeholder consortium convened by Friends of Cancer Research and aligned with FDA guidance on OS assessment, we developed a simulation-informed decision framework to aid interpretation of interim OS results, to support sponsor-CRO decision-making and governance. The approach simulates correlated PFS and OS endpoints under pre-specified 'truth' scenarios. Operating characteristics are estimated across different thresholds for harm, helping distinguish true treatment-related harm from artifacts of immature or evolving OS data, while quantifying the probability of detecting harm when it truly exists. The framework begins with the clinical question rather than the observed interim HR by first defining the anticipated OS effect, the degree of potential detriment considered clinically relevant, and plausible alternative survival trajectories. Candidate final decision rules may then be evaluated to determine whether the planned final event target can adequately characterize the relevant scenarios. Interim monitoring strategies can be quantitatively evaluated and trade-offs made to avoid misclassification of clinically relevant scenarios. We found that there is a lower limit to the amount of information (OS events) required at an interim OS analysis, to reliably make conclusion based on a resulting Hazard Ratio. Insufficient OS information is not necessarily solved by a more elaborate statistical rule. Simulation may instead show prospectively that the intended quantitative question cannot be answered reliably at the early endpoint readout at all. Recognizing that limitation before the data are observed can improve trial planning and reduce the risk of overinterpreting an unstable numerical estimate. We also describe practical ways to summarize multi-dimensional outputs through visual decision-support tools (e.g., heatmaps and scenario tables) that support transparent benefit-risk assessment. Attendees will leave with a pragmatic blueprint for planning interim OS analysis across a broad range of potential truths. By combining simulation, data visualization, and stakeholder-informed decision support, this work demonstrates how modern statistical leaders can translate uncertainty into actionable evidence, enabling more informed benefit-risk assessments and development decisions in oncology trials.
P02

Beyond Structured Data: A Large Language Model Pipeline for Extracting Osteoarthritis Severity and Pain from Clinical Notes

Dazheng Zhang
Read Abstract ▼
Structured claims/EMR data captured a lower TKA conversion rate, suggesting this discrepancy reflected unobserved OA severity and pain information not captured in structured fields. To address this, we developed a pipeline to extract OA severity, pain severity, and pain-at-night status from unstructured clinical notes. The pipeline followed four steps: (1) cohort definition via inclusion/exclusion criteria; (2) LLM-based natural language processing to determine patient-level phenotypes; (3) aggregation of note-level findings into a single patient-level evidence set; and (4) sampling of a subset of records for human expert review. Extraction was performed using an open-weight large language model deployed. A system prompt defined the model's role (clinical assistant specializing in OA), extraction task, evidence rules (only explicitly stated findings count), and a structured chain-of-thought process, with output returned as structured JSON. A note-specific user prompt then supplied each individual clinical note for extraction. Model performance was validated against clinician-adjudicated gold-standard labels on a randomly sampled set of notes, with extraction accuracy assessed separately for OA severity and pain.
P03
🏆 Reviewers' Choice · 2nd

A Calibrated Weibull Divide-and-Conquer Power Prior with Event-Driven Stratification for Survival Data

Dehua Bi
Read Abstract ▼
While external data can improve efficiency in randomized controlled trials (RCTs) with limited sample sizes, dynamic borrowing for time-to-event endpoints may lead to inflated Type I error when the external and current hazards differ. We propose a calibrated Weibull divide-and-conquer power prior for this setting. The method uses event-driven propensity score stratification to improve baseline comparability and applies borrowing-by-parts to the Weibull location and scale parameters. To further limit false-positive inflation, we introduce a pre-trial bootstrap calibration step that determines an additional penalty on the data-adaptive borrowing weights. The stratum-specific estimates are then combined to obtain overall treatment effects on the hazard ratio scale. Simulation studies and a real-data trial emulation based on the Netherlands Amyotrophic Lateral Sclerosis (ALS) Registry show that the calibrated procedure substantially reduces Type I error inflation relative to the uncalibrated approach while improving power and estimation accuracy relative to an unaugmented RCT. These results suggest that the proposed method provides a useful approach for incorporating external controls in confirmatory survival trials.
P04
🏆 People's Choice · 1st

Cross My TLF's and Dot My I's: An R Package for LLM-Assisted Cross-Checking of Clinical Trial Summary Tables

Diana Lam
Read Abstract ▼
Clinical study reports rely on hundreds of tables and summaries to communicate study results. Reviewing these outputs for consistency is an important but often time-consuming quality control activity. We developed tlfcrosschecks, an R package that uses large language models (LLMs) and meta data to automate cross-checks across clinical tables. The package reads clinical outputs, identifies related tables, and flags potential inconsistencies in areas such as patient counts, treatment groups, and adverse event summaries for further review. By helping reviewers focus on the outputs most likely to contain issues, this approach can reduce manual effort while maintaining human oversight. This work demonstrates a practical application of AI to support quality and efficiency in pharmaceutical reporting workflows
P05
🏆 People's Choice · 3rd

AI Assisted Code Review for Statistical Programming

Gianna Huang
Read Abstract ▼
Code review is one way of quality control in statistical programming, but it is often labor-intensive and requires significant reviewer effort to verify implementation against data mapping specifications. To improve review efficiency and consistency, we propose an AI-assisted code review framework that automatically evaluates SAS and R programs against approved mapping specifications. The proposed solution leverages the large language model (LLM) to interpret both the data mapping specification and the corresponding programming code, in order to traces the code execution path and assesses whether the implementation aligns with the documented specification. The output includes verification indicators, traceability summaries, and reviewer-friendly annotations that highlight potential discrepancies. Beyond specification compliance, the framework performs code-quality assessments to identify common programming risks, including hardcoded values, incomplete implementation of derivation logic, inefficient joins or merges, unnecessary data processing steps, and potential maintainability concerns. Findings are categorized by risk level, enabling reviewers to focus their attention on areas requiring expert judgment rather than routine verification tasks. The primary objective of this application is to augment reviewer productivity by providing a structured comparison between specifications and code, the solution can substantially reduce review time, improve consistency across projects, and increase the coverage of quality checks. In addition, the approach promotes standardization of programming practices and facilitates knowledge sharing among programmers with varying levels of experience. This AI-assisted review framework has the potential to enhance operational efficiency within statistical programming organizations by reducing manual effort while maintaining high quality standards.
P06

Modeling Longitudinal Cognition and Brain Volume Using a Mixed-Effects Model for Repeated Measures

Jasper Verma
Read Abstract ▼
Alzheimer's disease affects 13% of people aged 75-84, and 33% of people aged 85 and older. Alzheimer's disease and other forms of dementia can impair people's ability to function independently. The purpose of my research was to determine the extent of cognitive decline across time for patients diagnosed with dementia and to model the extent of cognitive decline across different subgroup categories of dementia patients. Subgroups include education level, age, and sex. The data I used was sourced from the OASIS dataset, which contains demographic information and MRI imaging statistics for patients with and without dementia. The file used in this analysis was oasis_longitudinal.csv, accessed through Kaggle. There were 64 dementia patients and 72 nondementia patients. Cognitive decline was measured through both MMSE (Mini-Mental State Examination) scores and nWBV (normalized whole brain volume), a measure from imagining scans. Patient MMSE scores, diagnosis, and nWBV are tracked across up to five visits (but only visits 1, 2, & 3 were used in the analysis due to small sample sizes). To do the analysis, an MMRM (Mixed Model with Repeated Measures) was used. The patients are divided into groups (Demented and Nondemented, and demographic subgroups of dementia patients) and the mean change in MMSE score and nWBV for each group (averaged across certain covariates) is tracked across visits. The decline in MMSE scores was significantly worse in dementia patients compared to nondementia patients. Both patients with dementia and patients without dementia showed significant declines in nWBV by both Visit 2 and Visit 3. There was no significant difference between change in nWBV for patients with dementia and change in nWBV for patients without dementia. There was no significant difference in decline in MMSE scores for dementia patients over age 75 as compared to patients at or under age 75. The decline in MMSE scores for dementia patients 14 or less years of education was significantly greater than the decline for patients with over 14 years of education. The decline in MMSE scores for female patients was significantly greater than the decline in MMSE scores for male patients.
P07
🏆 Reviewers' Choice · 1st

Towards Autonomous Real-World Evidence Generation: Agentic AI End-to-End Workflows

Jiying Zou
Read Abstract ▼
Real-World Evidence (RWE) generation traditionally relies on manual, time-intensive coding cycles. Legacy automation tools (such as dashboards and static reports) fail when subtle protocol tweaks across similar studies demand deep code refactoring rather than simple UI toggles. To solve this, we developed an agentic AI framework that orchestrates end-to-end RWE workflows - from protocol ingestion to final insight delivery - in a fast, reliable, and secure manner. Built upon domain-specific subject matter expertise, validated codebases, and automated quality control guardrails, the framework is grounded in core principles of trust & rigor, efficiency, transparency & auditability, and privacy. We run experiments to compare manual, non-agentic AI, and agentic AI workflow performance in terms of cost, speed, and reliability on classic data analysis tasks in the ophthalmology domain. By combining computational speed with expert human-in-the-loop oversight, this architecture elevates data scientists from routine script executors to strategic study architects, compressing delivery timelines and opening possibilities towards greater insight volume to drive faster, high-rigor impact for patients.
P08

Privacy-Preserving Local Unsupervised Signal Detection in Clinical Laboratory Data Using UMAP, HDBSCAN, and Isolation Forest

Kapil Sardiwal
Read Abstract ▼
Background Clinical trial teams need early, data-driven approaches for exploratory safety signal detection while maintaining confidentiality, governance, and control over sensitive data. Interest in AI-enabled exploration continues to grow, but many clinical data science use cases cannot rely on external platforms or services that require transfer of patient-level data. A local, privacy-preserving analytical workflow was developed to support unsupervised exploration of clinical laboratory data and identify patterns that may warrant clinical review. Objectives The objective was to build a reproducible local framework for exploratory safety signal discovery from ADLB-like laboratory data. The workflow aimed to combine nonlinear dimensionality reduction, density-based clustering, and anomaly detection to characterize population structure, identify atypical subject profiles, and generate interpretable outputs for clinical and cross-functional discussion without exposing sensitive trial data. Methods A proof-of-concept workflow was implemented locally using engineered laboratory features derived from subject-level and visit-level summaries, including observed values, change-from-baseline measures, and abnormality indicators. UMAP was applied to represent high-dimensional laboratory patterns in a lower-dimensional space while preserving local neighborhood structure. HDBSCAN was used to identify natural groupings without requiring a pre-specified number of clusters and to label sparse or atypical regions. Isolation Forest was applied to generate subject-level anomaly scores and support prioritization of records for review. Outputs included cluster summaries, subject-level anomaly tables, ranked laboratory parameters, treatment-stratified abnormality summaries, and visual displays designed for clinical interpretation. Results The local workflow successfully produced interpretable exploratory outputs for unsupervised review of laboratory data. The approach identified a manageable subset of subjects with atypical laboratory profiles, resolved distinct phenotype-like groupings, and highlighted laboratory parameters with recurrent abnormality patterns. Ranked outputs surfaced renal, metabolic, hematologic, and immune-related laboratory patterns that could support clinical review and hypothesis generation. Treatment-stratified summaries provided an additional lens for assessing whether selected laboratory patterns differed across groups. Importantly, the full workflow was executed within a governed local environment, avoiding external transfer of patient-level data. Discussion/Conclusion This work demonstrates that meaningful unsupervised signal detection can be performed within privacy-preserving local environments while maintaining analytical depth. The combined use of UMAP, HDBSCAN, and Isolation Forest provides complementary strengths: structure discovery, phenotype segmentation, and rare-pattern detection. As a practical data science method, the workflow supports early exploratory safety analytics, cross-functional review, and future signal triage under governance-first constraints. The approach is well suited for broader internal evaluation as a proof-of-concept framework for privacy-preserving clinical data science.
P09
🏆 People's Choice · 1st

TLG Agent - An LLM-Powered Framework for Automated Shell-to-Report Generation

Lucie Zhang
Read Abstract ▼
The development of clinical Tables, Listings, and Graphs (TLGs) from Statistical Analysis Plan (SAP) shells remains a labor-intensive and error-prone step in statistical programming workflows. Although the pharmaverse ecosystem has improved clinical reporting infrastructure, programmers still manually translate narrative shell specifications into executable R code, introducing inefficiencies, variability across programmers, and increased quality-control burden. To address this challenge, we developed TLG-Agent for clinical research, this is an LLM-driven framework that automates the conversion of shell specifications into reproducible R code. TLG-Agent parses shell content-including titles, footnotes, table structure, column definitions, and summary statistics-and converts these elements into structured prompts for a Large Language Model. The generated code is executed within a R Shiny to produce TLG outputs, allowing users to conveniently upload files and preview results in a transparent workflow.
P10
🏆 Reviewers' Choice · 3rd

Do Pretreatment Clinical Factors Add Predictive Value Beyond Tumor Mutational Burden for Immunotherapy Response in Metastatic Urothelial Carcinoma?

Mason Chang
Read Abstract ▼
Background: Checkpoint-inhibitor immunotherapy benefits only a minority of patients with metastatic urothelial carcinoma, yet non-responders still incur months of treatment, toxicity, and cost — checkpoint inhibitors can exceed $100,000 per year. Tumor mutational burden (TMB) is an established but incomplete biomarker of response. We hypothesized that routinely collected pretreatment features — spanning clinical status and the tumor immune microenvironment — carry predictive signals independent of TMB, and could sharpen patient selection without any new assay. Methods: Using the IMvigor210 cohort of atezolizumab-treated patients (n = 234; 61 responders, 26% response rate), we evaluated 22 candidate predictors spanning genomic/immune (log-TMB, TIDE, IFN-γ, MSI, CD274/PD-L1, CD8, T-cell exclusion, MDSC, CAF, TAM-M2, CTL, dysfunction) and clinical (metastatic site, TCGA subtype, ECOG performance status, IC/TC PD-L1, sex, tobacco, prior BCG, prior platinum, sample age) domains. Response was defined as complete/partial response (RECIST) versus stable/progressive disease. Two clinical predictors were dichotomized before modeling: metastatic site (LN-only/Missing vs. Liver/Visceral) and ECOG performance status (0 vs. 1–2). Variable selection was performed once on the full cohort, combining LASSO with stepwise AIC/BIC; selection stability was assessed via 200 subsampled LASSO replicates (75% resampling). The resulting three-variable model was fitted by standard logistic regression, and compared against a TMB-only model using likelihood-ratio test, AIC criteria, and DeLong tests across four cross-validation schemes (leave-one-out, single 5-fold, single 10-fold, repeated 5-fold). Because variable selection was not repeated within folds, reported AUCs reflect coefficient stability rather than selection stability. Results: Two genomic/clinical predictors formed a stable selection core — log-TMB (selected in 92.5% of subsamples) and metastatic site (85.0%) — with ECOG performance status added on clinical grounds despite a lower selection frequency (19.5%). Notably, none of the tumor immune-microenvironment features survived selection. All three predictors were significant in the multivariable model, with directions consistent with tumor biology: log-TMB OR 4.55 (95% CI 2.55–8.13, p < 0.0001), metastatic-site group OR 3.62 (1.79–7.31, p = 0.0003), and ECOG 0 OR 2.74 (1.35–5.58, p = 0.005). Adding the two clinical terms to TMB improved model fit (AIC 240.5 → 220.0; likelihood-ratio p < 0.0001) and raised out-of-sample discrimination by ≈ 0.08 AUC consistently across all schemes (LOO-CV 0.707 → 0.788; repeated 5-fold 0.713 → 0.788). The DeLong test confirmed a significant AUC gain in every scheme (p = 0.006–0.016), including all 20/20 repeated splits. Conclusions: Two variables already captured in routine clinical assessment — metastatic site and ECOG status — add reproducible, biologically coherent predictive value beyond TMB, with no additional testing burden; the tumor immune-microenvironment panel, by contrast, contributed no independent signal after selection. This is a promising single-cohort signal rather than a validated tool: because variable selection was performed outside the cross-validation loop, the reported AUCs may be optimistically biased, though the final three-variable model (~20 events per variable) is well-powered for stable coefficient estimation. Fully nested selection and external validation in an independent cohort are the necessary next steps before clinical use. Data source: IMvigor210 — Mariathasan S, Turley S, Nickles D, et al. TGFβ attenuates tumour response to PD-L1 blockade by contributing to exclusion of T cells. Nature 554, 544–548 (2018).
P11

Adaptive Trial Design - An AI-Enabled Workflow for Study Design Optimization and Stress-Testing

Matan Seidel
Read Abstract ▼
Choosing the right dose is an important milestone in the drug development process. On the one hand, sponsors are interested in making that decision as early as possible based on the lowest possible participant number in order to optimize speed to market and development costs. On the other hand, it is critical to select the dose with the optimal benefit and risk with high accuracy. Adaptive dose-finding trial designs theoretically enable optimal dose selection with as many participants as necessary based on emerging data. However, the expected operating characteristics of a dose-finding design are difficult to estimate and measure, making comparison of designs challenging and good parameter choices hard to find. We describe how, as part of a collaboration between Denali Therapeutics and PhaseV, we approached the design of dose-selection trials for novel disease modifying therapies as an iterative process: defining the dose-response scenario space and formal dose selection criteria, proposing alternative designs, simulating & measuring their operating characteristics, exploring the differences, and further refining the design & its goals. To make this process possible, we collaborated to extend PhaseV's Trial Optimizer, a configurable trial design and optimization platform. Within the platform, we implemented the capability to modify and optimize design parameters not usually available in off-the-shelf trial design software, including: the clinical utility function encoding the study's efficacy-toxicity trade-offs, weighting of the reference scenarios used to optimize and stress-test the design, and detailed assumptions about the desired operating characteristics. As a first step we investigated optimizing a standard multiple ascending dose design. The process began with making the trial's clinical priorities explicit by translating the target product profile into quantitative utility functions and acceptance criteria, then evaluating performance across weighted collections of investigator-defined and automatically generated scenarios using an efficient Monte Carlo engine. Each round of simulation surfaced new questions about how the design behaved, which investigators and methodologists examined together to decide what to adjust, test, or add next. With this evaluation framework in place, the team used a novel stochastic optimization algorithm to automatically explore the design parameter space, optimizing against the agreed clinical objective under realistic sample-size and duration constraints. This step revealed non-obvious parameter configurations and edge case performance that meaningfully improved the expected design performance over the original proposal. In a second step we extended the search space to include alternative designs with different interim decision time points and logic. Bayesian adaptive designs were implemented and optimized within the platform, applying isotonic regression to enforce monotonicity across doses in both efficacy and toxicity, with posterior probabilities then used to quantify certainty in the best dose and guide subsequent dose-selection decisions. Beyond these designs, the platform supported drafting and testing additional dose-selection algorithms, with an embedded AI assistant translating natural-language descriptions into efficient executable code. Competing strategies were then compared head-to-head using the same clinical objectives and evaluation framework. The outcome was not a single fixed design but a disciplined method for arriving at a better one: making clinical preferences explicit, stress-testing through large-scale simulation, optimizing, and comparing alternatives based on consistent assumptions and performance metrics until the design reflected both the study's scientific goals and its operational realities.
P12

From Feasible to Frictionless: Updates and Collaborative Milestones from the R Consortium Submissions Working Group

Ning Leng
Read Abstract ▼
Over the past several years, the conversation around open-source software in clinical trials has evolved from a fundamental question-Can R be used in regulatory submissions?-to a more consequential one: How can open-source submissions be made reproducible, transparent, and easy for regulators to review? Established under the R Consortium in 2021, the R Submissions Working Group brings together contributors across the pharmaceutical industry to collaborate directly with the FDA, address practical barriers in regulatory review, and establish open, reproducible precedents for the broader community. Through sustained industry-FDA dialogue and a series of public pilots, the group has explored regulatory submissions spanning static tables, listings, and figures; interactive Shiny applications; end-to-end ADaM workflows implemented in R; and portable applications using WebAssembly. These efforts also contributed to the FDA's August 2025 update of its eCTD file-format specifications to formally accommodate R package workflows, including formats such as .rds, .rdb, .rdx, .rdata, .md, .rd, .zip, and .html. The work is now expanding beyond demonstrating technical feasibility toward building a more scalable open-source regulatory ecosystem. Current initiatives include evaluating Dataset-JSON for analysis-data exchange, developing practical guidance and sponsor checklists for producing reviewable and reproducible code, and crowdsourcing reusable AI-agent skills and community benchmarks tailored to pharmaceutical development. Together, these efforts illustrate how cross-industry collaboration, open-source technology, and direct regulatory engagement can move the field from isolated demonstrations toward sustainable, community-driven standards for modern clinical trial submissions.
P13
🏆 People's Choice · 2nd

Combinations of Safety: Unifying Overall Survival Monitoring Strategies via '6 Choose 4'

Paloma Hauser
Read Abstract ▼
Advances in anticancer therapies have contributed significantly to decreasing death rates in certain disease and clinical settings. However, they have also made it difficult to power a clinical trial in these settings with overall survival (OS) as the primary efficacy endpoint. Therefore, two approaches have recently been proposed for the pre-specified analysis of OS as a safety endpoint (Fleming et al., 2024; Rodriguez et al., 2024). We provide in this poster (based on a recently published manuscript) a simple unifying framework that includes the aforementioned approaches-and a few others-as special cases. By highlighting each approach's focus, priority, tolerance for risk, and strengths or challenges for practical implementation, this framework can help facilitate discussions between stakeholders on "fit-for-purpose OS data collection and assessment of harm" (American Association for Cancer Research, 2023). We apply this framework to a real clinical trial in large B-cell lymphoma to illustrate its application and value.
P14

Adaptive Longitudinal K-Means Clustering of PSA Trajectories and Overall Survival in Metastatic Prostate Cancer

Phi Le
Read Abstract ▼
Background: Longitudinal biomarkers provide evolving information about disease course, yet identifying clinically meaningful patterns from accumulating repeated measurements remains challenging. In metastatic prostate cancer, prostate-specific antigen (PSA) is measured repeatedly during treatment, and patients may exhibit heterogeneous patterns of change over time. Approaches that accommodate accumulating longitudinal information and evolving patient group membership may better characterize these dynamic patterns. We aimed to identify longitudinal PSA trajectory patterns and evaluate their association with overall survival (OS). Methods: We developed an adaptive longitudinal k-means clustering approach in which cluster membership was updated as longitudinal measurements accumulated. At baseline, patients were clustered based on baseline log-transformed PSA. At each subsequent monthly time point, k-means clustering was applied to an expanding window of log-PSA measurements from baseline through that time point, guided by the preceding cluster solution and allowing cluster membership to evolve as new longitudinal information became available. Cluster assignments, after incorporating the month 6 PSA measurement, defined the final longitudinal cluster membership, and cluster centroids were used to characterize the mean longitudinal log-PSA trajectory within each final cluster. The approach was applied to pooled data from six phase III clinical trials (TITAN, LATITUDE, COU-AA-301, COU-AA-302, ACIS, and MAGNITUDE) and one open-label phase II trial (GALAHAD), comprising 5,551 patients with metastatic castration-resistant prostate cancer (mCRPC; n=3,321, 60%) or metastatic hormone-sensitive prostate cancer (mHSPC; n=2,230, 40%). PSA measurements at seven time points from baseline through month 6 were analyzed, with missing values imputed using multiple imputation by chained equations (MICE). OS from randomization was estimated using Kaplan-Meier methods and compared across final trajectory clusters using the log-rank test. Analyses were also conducted separately for mCRPC and mHSPC. Results: Three distinct longitudinal log-PSA patterns were identified. Cluster 1 was characterized by persistently high log-PSA with minimal decline; Cluster 2 showed an early decline from an intermediate baseline level followed by relative stabilization; and Cluster 3 showed a steep early decline from a lower baseline level followed by continued gradual decline. Median OS was 16.23, 36.44, and 62.52 months for Clusters 1, 2, and 3, respectively (log-rank P<0.001). Similar separation in OS across trajectory clusters was observed within the mCRPC and mHSPC subgroups, although OS was longer overall in mHSPC. Conclusions: Adaptive longitudinal k-means clustering identified distinct PSA trajectory patterns associated with substantially different OS among patients with metastatic prostate cancer. Persistently high PSA was associated with shorter survival, whereas an early and sustained decline was associated with longer survival. These findings support further evaluation of adaptive longitudinal clustering as an approach for dynamic patient stratification.
P15

A Mallows-Centered Latent Ordering Model for the Partial Ordering Continual Reassessment Method in Drug-Combination Trials

Qianyu Dong
Read Abstract ▼
The Partial Ordering Continual Reassessment Method (POCRM) is a model-based design for Phase I clinical trials when the toxicity ordering of treatment combinations is only partially known. POCRM extends the CRM skeleton power model framework by considering multiple plausible toxicity orderings of the treatment combinations. Accurately representing uncertainty across plausible toxicity orderings is important for guiding treatment allocation during the trial. We propose a cohort-specific latent-ordering extension of POCRM that treats ordering weights as unknown parameters. The weights follow a Dirichlet model centered on a prespecified distribution over candidate orderings. A Mallows model on Kendall distance allows greater prior weight to be assigned to orderings closer to a reference ordering. Each cohort is assigned a latent ordering according to the current ordering weights, and each ordering has its own CRM power parameter. Posterior inference is using a Metropolis-within-Gibbs algorithm. We consider two treatment recommendation strategies: Bayesian model averaging, which averages toxicity estimates across orderings, and a maximum-a-posteriori ordering rule, which first selects the ordering with the largest posterior mean weight. The proposed framework provides a flexible approach for representing ordering uncertainty while incorporating prior scientific information. Simulation studies evaluate operating characteristics, sensitivity to prior specification, and dose-allocation behavior against existing POCRM approaches.
P16

Prompt Design for Multi-Criteria Variable Extraction: A Case Study of Tumor Grade in Neuroendocrine Neoplasms Using Large Language Models

Qiyun Li
Read Abstract ▼
Background: Manual abstraction of clinical variables from pathology reports is labor-intensive, particularly for variables such as tumor grade that may require integrating multiple sources of information, including explicitly stated grade, Ki-67 proliferative index, and descriptive grade terminology. Large language models (LLMs) can automate such extraction, but it remains unclear whether prompts that integrate multiple prespecified grading criteria outperform simpler, single-criterion prompts. We evaluated how prompt design affects LLM extraction using tumor grade in neuroendocrine neoplasms (NENs) as a test case. Methods: We used OpenAI's GPT-4o via a secure, IRB-approved platform to extract tumor grade (G1-G3 or Unknown) from 364 NEN pathology reports. Manual abstraction served as the reference standard. For this analysis, three prespecified sources of grading information were represented in the prompts: explicitly stated grade; grade inferred from the Ki-67 proliferative index; and descriptive grade terminology (low grade -> G1, intermediate grade -> G2, high grade -> G3). Four LLM prompt strategies isolating or combining these were evaluated: explicit-grade-only, Ki-67-inference-only, descriptive-terminology-only, and a combined prompt that sequentially applied all three criteria in the prespecified order (explicit grade -> Ki-67 -> descriptive terminology), returning Unknown if none yielded a grade. Agreement with the manually abstracted reference was evaluated using accuracy and Cohen's kappa. We also evaluated performance among cases with known reference grades. Results: Overall, the descriptive-terminology-only prompt had the highest agreement with the manual abstraction (accuracy= 85.2%, kappa = 0.780), followed by the combined hierarchical prompt (83.0%, 0.735), the explicit-only prompt (81.3%, 0.732), and the Ki-67-inference-only prompt (79.4%, 0.708). Of 364 reports, 31 (8.5%) had an Unknown grade in the manual reference standard. The proportion classified as Unknown differed across prompting strategies: 19.2% for explicit-only, 15.9% for Ki-67-only, 14.3% for descriptive-only, and 3.0% for the combined prompt. In a sensitivity analysis restricted to cases with a known grade in the manual reference standard, the ranking of the four prompting strategies was changed, with the combined prompt achieving the highest accuracy (89.8%), followed by descriptive-only (85.9%), explicit-only (80.2%), and Ki-67-only (79.3%). Conclusion: Prompt architecture was associated with both tumor-grade classification and the tendency to return Unknown. Although the descriptive-only prompt achieved the highest overall agreement with manual abstraction, the combined hierarchical prompt performed best among cases with known reference grades, indicating that aggregate accuracy may obscure differences between grade classification and abstention behavior. Evaluation of multi-criteria LLM extraction should therefore assess these components separately. Future work will capture the criterion and the supporting source text for each extracted value, incorporate additional contextual information, such as tumor differentiation, and evaluate whether structured representations of hierarchical rules improve adherence to expert abstraction logic.
P17

Predicting Viral Infection From TCR Repertoires Using Sequence Similarity and Graph-Based Features

Richard Adelstein
Read Abstract ▼
T-cell receptor (TCR) repertoire data provide a snapshot of an individual's immune system. Viral infection may alter an individual's TCR repertoire in ways that can be detected by statistical models. Many existing methods of inferring viral infection and exposure from TCR repertoire data focus on the presence or abundance of specific individual public clonotypes; however, these methods may end up leaving out information from similar, but less common TCR clonotypes. We investigate several methods of incorporating information from 'neighborhoods' of important TCR receptors into repertoire-level predictions of viral infection. First, we identify a small group of TCR clonotypes which are strongly associated with viral infection. Then with these clonotypes as anchors, we build features describing their neighborhoods, using various notions of similarity such as edit distance, prelearned sequence embedding distance, and graph embeddings. We evaluate to see whether these features improve predictive performance of downstream classifiers in comparison to features based solely on presence/abundance of individual clonotypes. Ideally, this framework can be adapted to a variety of TCR repertoire-level prediction tasks.
P18

Bayesian Borrowing from Historical Controls under Local Non-exchangeability

Shike Xu
Read Abstract ▼
Borrowing historical or external controls can make a randomized trial more efficient, but only when those controls are comparable to the current ones. Bayesian dynamic borrowing methods, including the power prior, the commensurate prior, and the self-adapting mixture prior, discount the external data through a single global weight for the entire dataset, which limits their ability to distinguish compatible external patients from incompatible ones. Covariate-based methods such as propensity score matching act at the patient level and balance the baseline covariate distribution, which addresses exchangeability only when source differences are fully explained by the measured covariates. If the outcome relates to X differently across sources, covariate balance alone may not reveal it. Two outcome models can agree in one region of the covariate space and differ in another, a pattern we call local non-exchangeability, and neither a global weight nor covariate balance is designed to detect it. We bring the conditional distribution of the outcome into the matching step. First, using a prespecified ordering covariate, we fit segmentation regressions to find cutpoints that partition the external and current controls into subgroups with an approximately homogeneous Y given X relationship. Second, each external subgroup is compared with the corresponding current controls, and that exchangeability assessment sets its borrowing weight through an empirical Bayes power prior or a commensurate prior. Subgroup contributions are then combined into an estimate of the current-trial average treatment effect. In simulations, the estimated weights recover the underlying local structure, borrowing fully, partially, or almost not at all as subgroups were exchangeable, mildly different, or non-exchangeable. Bias remained low and the method gave the largest RMSE reduction compared to global borrowing and propensity score matching methods.