Large-scale EMR-based machine learning for early risk stratification of colorectal cancer.

Abstract

e15513 Background: Early detection of colorectal cancer (CRC) substantially improves outcomes; however, age- and symptom-based screening strategies fail to identify many high-risk individuals. Machine learning (ML) applied to longitudinal electronic medical records (EMR) may enable earlier, data-driven risk stratification beyond traditional risk factors. Methods: We analyzed a de-identified EMR event-stream dataset comprising 61,062,985 records from 300,000 patients. A case–control CRC cohort was constructed including 500 incident CRC cases (43,123 records) and 299,500 cancer-free controls with ≥3 years follow-up beyond the last observation. CRC onset was defined as the multidisciplinary tumor board diagnosis date. To reduce diagnostic work-up leakage, events within 6 months prior to onset were excluded.Patient histories were featurized over a primary prediction horizon of −36 to −6 months, summarized across four pre-diagnostic intervals (−6 to −12, −12 to −24, −24 to −36, and > −36 months). Models (random forest, LightGBM, histogram-based gradient boosting) were trained using patient-level splits, class rebalancing (SMOTE), and combined via a soft-voting ensemble.Generalizability was assessed in an independent, non-overlapping validation cohort (~100,000 patients) without restrictions on comorbidities and with CRC prevalence reflecting the general population. Results: In internal testing, the ensemble achieved accuracy 80.0%, sensitivity 72.2%, specificity 80.1%, NPV 95.0%, F1-score 75.9%, and ROC AUC 0.8175. Performance was preserved in the population-representative validation cohort (accuracy 79.0%, sensitivity 70.3%, specificity 79.0%, NPV 99.0%, F1-score 74.4%, ROC AUC 0.8217).Key contributors included age, healthcare utilization patterns across multiple specialties, routine laboratory parameters, and vital signs, rather than cancer-specific markers. Window-specific models for shorter pre-diagnostic intervals are under evaluation. Conclusions: Machine learning applied to large-scale longitudinal EMR data enables early CRC risk stratification up to three years before diagnosis using nonspecific, routinely collected features. Validation in an independent non-overlapping cohort supports feasibility as a phase-1, population-scale triage approach. Limitations include potential confounding by healthcare utilization intensity. Prospective evaluation of clinical yield (including PPV), sensitivity analyses, decision-curve utility, and validation across healthcare systems are planned.