AI for Precision Oncology Evidence: Case Study Research From British Columbia, Canada
Author(s)
Deirdre Weymann, MA1, Emanuel Krebs, MA2, Dean Regier, BA, MA, PhD3.
1Regulatory Science Lab, Simon Fraser University / BC Cancer, Burnaby, BC, Canada, 2Regulatory Science Lab, BC Cancer, Vancouver, BC, Canada, 3Regulatory Science Lab, University of British Columbia / BC Cancer, Vancouver, BC, Canada.
1Regulatory Science Lab, Simon Fraser University / BC Cancer, Burnaby, BC, Canada, 2Regulatory Science Lab, BC Cancer, Vancouver, BC, Canada, 3Regulatory Science Lab, University of British Columbia / BC Cancer, Vancouver, BC, Canada.
OBJECTIVES: Globally, patient access to precision oncology is variable. Healthcare systems are ill equipped to rapidly integrate research data with real-world health and equity information to generate evidence that supports precision oncology access. We sought to determine whether AI approaches can de-silo real-world data confined within electronic health records (EHRs) and enable comparative evidence generation for precision oncology.
METHODS: We undertook two AI case studies in British Columbia, Canada. Our first study validated natural language processing (NLP) and large language models (LLMs) for automating real-world data extraction from EHRs for adult patients enrolled in a precision oncology program. For 24 clinical features of interest, we evaluated accuracy of AI-extracted data compared to manually curated ground truths. Our second case study determined how de-siloed data and machine learning methods can support evaluations of emerging targeted therapies. Focusing on entrectinib, a TRK-inhibitor, we fused single-arm phase I/II trial data with cross-jurisdictional real-world data from British Columbia and the United States (Flatiron Health). We applied genetic algorithm matching and clone-censor weighting in target trial emulations of entrectinib versus standard care, estimating comparative survival over one-year.
RESULTS: For our first case study, we obtained 113,024 EHR documents, consisting of 194 distinct document types, for 211 patients meeting eligibility criteria. An LLM engine automated extraction of all clinical features of interest, achieving accuracy of ≥85% per feature. Our second case study identified 210 eligible patients with a rare biomarker, 55 (26%) of whom received entrectinib. Target trial emulations identified non-significant signal of survival benefit from entrectinib compared to standard care in a tumour-agnostic setting (aHR: 0.43, 95% CI: 0.10, 1.79 based on genetic matching; and aHR: 0.66, 95% CI: 0.33, 1.30 based on clone-censor-weighting). Tumour-specific effectiveness was highly uncertain.
CONCLUSIONS: AI-augmented data and evidence can inform life-cycle decision-making and support patient access to precision oncology.
METHODS: We undertook two AI case studies in British Columbia, Canada. Our first study validated natural language processing (NLP) and large language models (LLMs) for automating real-world data extraction from EHRs for adult patients enrolled in a precision oncology program. For 24 clinical features of interest, we evaluated accuracy of AI-extracted data compared to manually curated ground truths. Our second case study determined how de-siloed data and machine learning methods can support evaluations of emerging targeted therapies. Focusing on entrectinib, a TRK-inhibitor, we fused single-arm phase I/II trial data with cross-jurisdictional real-world data from British Columbia and the United States (Flatiron Health). We applied genetic algorithm matching and clone-censor weighting in target trial emulations of entrectinib versus standard care, estimating comparative survival over one-year.
RESULTS: For our first case study, we obtained 113,024 EHR documents, consisting of 194 distinct document types, for 211 patients meeting eligibility criteria. An LLM engine automated extraction of all clinical features of interest, achieving accuracy of ≥85% per feature. Our second case study identified 210 eligible patients with a rare biomarker, 55 (26%) of whom received entrectinib. Target trial emulations identified non-significant signal of survival benefit from entrectinib compared to standard care in a tumour-agnostic setting (aHR: 0.43, 95% CI: 0.10, 1.79 based on genetic matching; and aHR: 0.66, 95% CI: 0.33, 1.30 based on clone-censor-weighting). Tumour-specific effectiveness was highly uncertain.
CONCLUSIONS: AI-augmented data and evidence can inform life-cycle decision-making and support patient access to precision oncology.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
RWD13
Topic
Health Policy & Regulatory, Methodological & Statistical Research, Real World Data & Information Systems
Disease
Oncology, Personalized & Precision Medicine