Presentation (CTI)

OBJECTIVES: This study describes the reliability, completeness, and internal validity of a novel machine learning (ML)-generated real-world response (rwR) approach.
METHODS: This study used the nationwide, Flatiron Health electronic health record (EHR)-derived, de-identified database. A deep learning-based, natural language processing model extracted clinicians’ documentation of changes in disease burden (complete response [CR], partial response [PR], stable disease, progressive disease, or unknown) at imaging timepoints. Data from 18 treatment and/or biomarker-defined cohorts across 7 solid tumors were used to train the model and test the correlation between human-abstracted and ML-extracted real-world response rate (rwRR). In 15 cohorts of common solid tumors, the proportion of treated patients with at least 1 assessment and the time to first, second, third, and median number of assessments for first (1L), second (2L), and third lines (3L) of therapies, were examined. Additionally, real-world overall survival (rwOS) was compared for responders (ever achieved CR or PR) versus non-responders (never achieved CR or PR) for the most frequent regimens in 1L to 3L for each disease (with ≥30 patients).
RESULTS: Within the test cohort (n = 4047), the correlation between human-abstracted and ML-extracted rwRR was r = 0.86. The solid tumor cohorts included 3406-129 807 treated patients. 57.8%-80.6% of patients had at least 1 assessment, with a median of 1-3 assessments within 1L, 2L, and 3L. Median times to first, second, and third assessments for 1L-3L were 1.9-4.4, 3.8-8.7, and 5.7-12.9 months, respectively. Across all most frequent 1L-3L regimens for each disease, responders for each cohort had significantly longer survival compared to non-responders (P < .05).
CONCLUSIONS: This study establishes the performance and validation of a novel ML approach for capturing rwR data from EHRs; supporting the efficient and reliable generation of valuable outcome data across large cohorts.