Toward Equitable Precision Oncology: Real-World Validation of an AI-Powered Trial Matching Platform
Yan Leyfman, Arturo Loaiza-Bonilla, Ertugrul Tuysuz, Selin Kurnaz, Oz Huner, Juan P.N. Menza, Serkan Yerdan, Cagatay Cagatay, Zara Baloch, Adeena MusheerPURPOSE
Limited oncology workforce capacity, fragmented electronic health records (EHRs), and labor-intensive chart abstraction contribute substantially to under-enrollment in cancer clinical trials worldwide, particularly in resource-constrained settings. We evaluated a GPT-4o–based artificial intelligence platform designed to automatically extract and standardize oncology data from unstructured EHRs using the Minimal Common Oncology Data Elements (mCODE) framework to improve scalable and interoperable clinical trial screening.
METHODS
A fine-tuned oncology-specific GPT-4o pipeline integrated within the Massive Bio platform was developed to process free-text oncology documentation and structured EHR data. A random cohort of 102 patients was selected from a real-world oncology database of 3,800 cases representing diverse malignancies and genomic profiles. Extracted variables included tumor type, disease stage, metastatic status, relapse status, resectability, and biomarker results derived from next-generation sequencing and immunohistochemistry. Outputs were mapped to mCODE 3.0 profiles, including Cancer Disease Status, Tumor Profiles, and Genomic Profiles. AI-generated outputs were validated against expert manual chart abstraction. Primary endpoints included variable-level accuracy using F1-based performance metrics. Secondary endpoints included completeness and interoperability of mCODE mapping.
RESULTS
The AI platform demonstrated high performance across multiple oncology domains. Accuracy was highest for tumor type (98%) and extent of disease classification (90%). Tumor stage achieved 86% accuracy, with discrepancies primarily involving sub-stage assignment. Relapse status and resectability achieved 77% and 69% accuracy, respectively, reflecting inconsistent surgical and longitudinal clinical documentation. Genomic biomarker extraction achieved 78% accuracy overall, with strongest performance for common pathogenic variants including BRCA1/2 and TP53. Demographic variables demonstrated 100% concordance. The system successfully populated all evaluated mCODE profiles with high completeness and interoperability. Compared with conventional manual abstraction, the platform substantially reduced labor intensity and time requirements for trial screening workflows.
CONCLUSION
This study demonstrates the feasibility of combining GPT-4o–based oncology abstraction with mCODE interoperability standards to support scalable and standardized clinical trial matching. By reducing dependence on manual chart review, AI-driven oncology infrastructure may help expand equitable access to precision oncology trials, particularly in settings with limited oncology workforce capacity. These findings support further development of deployable, interoperable AI systems capable of accelerating evidence generation and improving global access to cancer clinical research.