OBJECTIVES: This study evaluates real-world response rate (rwRR) in clinical trial-aligned cohorts of patients with lung, colon and breast cancer using machine learning.

METHODS: This study used the nationwide Flatiron Health EHR-derived de-identified database. Machine learning was leveraged to extract clinician’s documentation of change in disease burden at imaging assessments for response determination. Four real-world cohorts were generated to align with the control arms of the ALEX, KEYNOTE-021, KEYNOTE-177, and PALOMA-2 trials. Cohorts included patients within the same enrollment windows as the respective trials, and were aligned by implementing trial inclusion/exclusion criteria where feasible and via inverse odds weighting on available baseline characteristics (e.g., age, gender, race, ECOG, smoking status, stage, treatment, number/sites of metastases). rwRRs were calculated with and without confirmation, before and after weighting, and qualitatively compared with each trial’s objective response rate (ORR).

RESULTS: In total 106, 384, 1251, 268 patients were included in the trial-aligned real world cohorts for ALEX, KEYNOTE-021, KEYNOTE-177, PALOMA-2 respectively. Their weighted rwRRs(%) with 95% CI, with real-world confirmation when required per trial protocol, vs ORR, were 37.9 (27.1, 50.1) vs 75.5 (67.8, 82.1), 31.3 (26.4, 36.7) vs 29 (18, 41), 37.0 (34.1, 39.9) vs 33.1 (25.8, 41.1), 19.3 (11.2, 31.1) vs 34.7 (28.4, 41.3), and absolute differences were -37.6%, 2.3%, 3.9%, -15.4%.

CONCLUSIONS: While ORRs and rwRRs were similar for KEYNOTE-021/-177, differences were more pronounced in other trials, likely reflecting less ability of matching trial characteristics in real-world data and the differences in real-world care where therapy adherence, assessment cadence and other factors may confound/impact results. It also highlights, depending on the cohort, the importance of prioritizing more thorough capture of key real-world characteristics or utilizing patient-level trial data for more robust matching.