Rule-Based Identification of a Reliable Real-World Cancer Recurrence Endpoint.
TL;DR
This scalable method enables derivation of real-world endpoints from EHR-based data and may be used to characterize outcomes for patients treated outside of prospective clinical trials and may inform the creation of synthetic control cohorts to support regulatory approvals for new drugs in rare tumor or biomarker defined populations.
Abstract
Background
Recurrence is a key oncologic endpoint but is difficult to automatically capture from electronic health records (EHR).
Methods
We evaluated rule-based algorithms to detect recurrence and its timing using a publicly available clinico-genomic database of patients with breast, colorectal, non-small cell lung, or pancreatic cancer. Algorithms evaluated varying anchor dates, defined as the time at which patients were assumed eligible to recur for the purposes of the algorithm, including diagnosis and four-, six-, and twelve-months post-diagnosis, as well as varying criteria for subsequent evidence of cancer from radiology, pathology, medical oncology assessments, or cancer-directed regimen initiation. Algorithm-derived recurrence results were compared with manually curated institutional data. Performance was measured by sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and concordance of recurrence timing.
Results
The best algorithms, anchored at six months post-diagnosis and requiring two reports demonstrated high sensitivity (≥0.95), PPV (≥0.81), and NPV (≥0.84), reasonable specificity (≥0.75), and estimated recurrence within a median of one month of institutional reference data.
Conclusions
This scalable method enables derivation of real-world endpoints from EHR-based data. IMPACT The algorithm may be used to characterize outcomes for patients treated outside of prospective clinical trials and may inform the creation of synthetic control cohorts to support regulatory approvals for new drugs in rare tumor or biomarker defined populations.