A dynamic time-lagged machine learning approach to early colorectal cancer detection in primary care using electronic health records

Abstract Early detection of colorectal cancer (CRC) remains a major unmet need, as current diagnostic pathways rely on symptoms which often appear only at advanced stages. Longitudinal electronic health records (EHRs) provide the means to identify earlier signals, and are well suited to machine learning approaches capable of identifying complex temporal patterns associated with CRC. We present DetectCRC, an XGBoost-based model for early detection of CRC using real-world EHR data. Using 13.6 million QResearch primary care records, the model is developed and evaluated in a matched nested case-control study (22,406 CRC cases; 559,849 controls), incorporating demographics, clinical risk factors, and two years of longitudinal symptom and blood test data using time-lagged features in 6-month advancing windows prior to diagnosis. Longitudinal analyses show that symptom signals are largely confined to the 6 months preceding diagnosis, whereas changes in blood test results, particularly full blood count markers, are detectable up to two years prior. Although multiple biomarkers are evaluated, optimal performance is achieved using routinely available full blood count tests alongside symptoms and risk factors. DetectCRC achieves a C-index of 0.874 in a held-out regional validation set and maintains strong performance on sparse data, demonstrating the potential of routinely collected blood tests for earlier CRC detection.