A machine learning–enhanced analysis of temporal trends in biomarker-based eligibility restrictions in oncology clinical trials.

11019 Background: The adoption of biomarker-based eligibility criteria reflects the evolution of precision oncology, but temporal trends and predictors of biomarker restrictions are unclear. Understanding these patterns is essential for optimizing trial design and ensuring equitable patient access to investigational therapies. We used machine learning and natural language processing to analyze biomarker restriction trends across oncology trials. Methods: We analyzed all 54,985 oncology interventional trials from ClinicalTrials.gov from 2010-2024. Biomarker restrictions were identified using curated term lists plus NLP-based classification (TF-IDF vectorization with logistic regression). We developed complementary machine learning modules for emerging biomarker discovery using named entity recognition, requirement type classification (inclusion/exclusion/stratification), and eligibility complexity scoring via text feature engineering. Statistical analyses included logistic regression with Wilson score confidence intervals and Benjamini-Hochberg FDR correction. Results: Biomarker-restricted trials increased from 20.8% (2010) to 31.2% (2024), representing a 1.5-fold increase over 15 years (OR per year: 1.046, 95%CI: 1.041-1.050, p<0.001). Industry sponsorship (OR: 1.88, 95%CI: 1.80-1.97) and NIH sponsorship (OR: 1.89, 95%CI: 1.67-2.13) independently predicted higher biomarker restrictions versus academic sponsors (all p<0.001). US-based trials showed 40% higher odds of restrictions (OR: 1.40, 95%CI: 1.35-1.46). Phase II trials demonstrated the highest rates (40.9%). Among specific biomarkers, PD-1 showed the steepest adoption trajectory (0.08% to 10.59%), followed by EGFR (5.07% to 8.84%) and HER2 (5.16% to 8.17%). Our NLP-based emerging biomarker discovery identified CTLA-4, PSMA, CD19, CD137, and LAG-3 as rapidly emerging eligibility biomarkers reflecting next-generation immunotherapy targets. Cancer-specific analyses revealed dramatic increases in lung (22.8% to 50.5%) and gastric cancers (14.6% to 45.5%), while prostate cancer showed declining rates (60.5% to 52.3%). Our complexity scoring algorithm demonstrated that biomarker-restricted trials have significantly higher eligibility complexity (mean: 50.0 vs 31.0, p<0.001). All findings remained robust across multiple sensitivity analyses. Conclusions: Biomarker-based eligibility restrictions in oncology trials have increased substantially over 15 years, driven primarily by immune checkpoint and targeted therapy biomarkers. Industry-sponsored and later-phase trials show the highest adoption rates. These scalable ML findings inform precision medicine trial design and highlight considerations for patient access and trial generalizability.