2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Technology and Procedures, Assays, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Clinical Procedures, Computational Biology, Emerging Technologies, Gene Editing, Profiling, Treatment Considerations, Drug Development, Therapy Sequence, Biological Therapies, Transplantation, (CAR)-T Cell Therapies, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Pathogenesis, CHIP, Genomics, Hematopoiesis, Immune Mechanism, Metabolism, Microbiome, Molecular Biology, Multi-Systemic Interactions, Clinical Practice (Health Services and Quality), Quality Improvement, Systems-Based Hematology, Diseases, Bone Marrow Failure Syndromes, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Measurable Residual Disease, ALL, CLL, LGL, Hodgkin Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Elderly
PAPER

4339 A tiered risk stratified artificial intelligence framework for clinical trial data monitoring: Prototyping AI tools for the phase II BRAVE study

Kyle B Jones, DDS, PhD, Jinnie Ko, PhD2, Yong Mun, PhD3, Vivek S. Chopra, PhD, MSc4, Darshna Desai, MS5, Dana Farinella, PharmD6, Richard H. Ko, MD, MHS, MS7, Mina Shahkarami, PharmD8
1 Genentech, Inc., South San Francisco, California, United States
2 Genenetch, Inc., South San Francisco, California, United States
3 Genentech, Inc., South San Francisco, California, United States
4 Genetech, Inc., South San Francisco, California, United States
5 Genentech, Inc., South San Francisco, California, United States
6 Genentech, Inc., South San Francisco, California, United States
7 Genetech, Inc., South San Francisco, California, United States
8 Genentech, Inc., South San Francisco, California, United States
Background: Medical monitoring in clinical trials is bottlenecked by manual data processing, leaving teams reliant on labor intensive text reviews and overly complex eCRFs designed primarily for audit compliance. Workload analyses highlight an urgent need for automation: One clinical trial reported generating over 21,100 data queries, 36% (7,560) of which required manual execution. At a cost of ~$160 per query, this added >$1.2 million in operational costs (Bradford, J Clin Stud, 2021). While Large Language Models (LLMs) offer strong potential to streamline these workflows, trial operations still need practical, clear guidance on integrating artificial intelligence (AI) into routine practice through risk-stratified monitoring and human oversight. We propose a three tiered risk framework governing AI
utilization for trial data quality assurance and evaluate its pilot deployment in the Phase II BRAVE (BTKi Responders to Achieve deep remission with VEnetoclax) study (NCT06524375) in chronic lymphocytic leukemia.

Methods: The risk-stratified framework categorizes deployment based on clinical impact and required oversight: 
● Tier 1 (Low Risk, Internal Automation): High autonomy for routine structured tasks (e.g., extracting data from clonoSEQ MRD reports), governed by periodic audits instead of line-by-line manual review
● Tier 2 (Medium Risk, Internal Evaluation): AI or algorithm assisted synthesis of complex data, requiring mandatory Medical Monitor adjudication before outputs inform internal decisions (e.g., iwCLL response evaluation integrating clinical exams, labs, and imaging)
● Tier 3 (High Risk, External Action): Outputs reaching external stakeholders (sites, regulators) requiring formal Medical, Legal, and Regulatory human review Using Claude Code (CC) via natural language prompting without formal software engineering, we prototyped two proof of concept automated pipelines for the BRAVE study: a Tier 1 automated PDF data extractor and a Tier 2 Streamlit based iwCLL response auditing dashboard. Tier 3 prototyping will be piloted in a later phase.

Results: In Tier 1 testing, the CC generated data extractor processed 196 screening and end of therapy clonoSEQ MRD reports (containing a combined 67 unique data elements) in ~20 seconds. Stratified random sampling of 15 baseline and 15 MRD tracking reports confirmed 100% extraction accuracy, saving ~32 hours of manual transcription and eliminating data entry errors.

In Tier 2 proof of concept testing, the CC generated algorithm exceeds standard end of study binary yes/no checks by enabling real time, interactive data review with contextual discrepancy explanations. The algorithm independently applies iwCLL criteria across longitudinal eCRFs, flagging discordant site versus calculated responses with structured rationales linked to specific data fields. Concurrently, 23 automated data quality checks across 15 categories identify missing values and eCRF inconsistencies. Each finding is graded across 7 clinical impact tiers, from "could trigger progressive disease" to "no impact on response". Findings populate in real time within an interactive Streamlit dashboard where reviewers access subject cards displaying iwCLL response, the discrepancy narrative, prioritized data quality findings, and autogenerated plain language queries to resolve critical data gaps. Because the BRAVE study data remains immature, we validated algorithm accuracy using a synthetic cohort of 100 subjects designed to encompass all iwCLL response and data quality categories, including edge cases from real world clinical trial data. The algorithm correctly classified 100% of synthetic response scenarios and detected all quality flag categories. Deployment on the active BRAVE study is underway, and prospective operational impact metrics will be presented.

Conclusions: This tiered risk framework provides a blueprint for safely deploying AI and LLM generated algorithms in clinical trial operations. This approach accelerates review, lowers costs, and maintains rigorous human oversight. These algorithms also improve upon standard end of study binary checks by synthesizing multidimensional data to identify actionable discrepancies in real time. This structured output enables future explainable machine learning analyses (e.g., XGBoost with SHAP) for biomarker discovery. Ultimately, this hybrid operational model balances automation and AI efficiency with human expertise to deliver scalable, reliable trial monitoring.
Kyle B Jones, DDS, PhD: Genentech, Inc. - Current equity holder in publicly-traded company; Genentech, Inc. - Stock (publicly traded); Genentech, Inc. - Current Employment