Machine Learning Improves Colorectal Cancer Recurrence Prediction

MACHINE learning models predicted overall survival and site-specific recurrence in patients with colorectal cancer liver metastases (CRLM) following curative-intent hepatectomy, with dynamic models providing stronger discrimination than baseline models.

The retrospective study included 730 patients with CRLM who underwent curative-intent hepatectomy at Johns Hopkins Hospital between 2000–2024. Median follow-up was 10.9 years, and more than 63% of patients experienced recurrence.

Researchers developed and validated models to predict recurrence in the liver, lung, lymph nodes and peritoneum, alongside overall survival. The models incorporated clinical, pathologic and molecular features and used baseline and dynamic prediction frameworks.

Dynamic Models Improve Risk Prediction

Baseline models achieved concordance indices (C-indices) ranging from 0.54–0.67. Dynamic models incorporated surveillance information collected during follow-up, including adjuvant therapy and recurrence at other anatomical sites.

This approach improved discrimination, with C-indices of 0.67–0.78 and time-dependent area under the curve (AUC) values of 0.68–0.90. Calibration was also assessed using Brier scores, which ranged from 0.03–0.19.

The data indicated that recurrence risk could be updated as patients progressed through surveillance rather than relying solely on information available at the time of surgery. This was relevant because recurrence patterns varied by anatomical site.

Implications for Personalised Surveillance

The researchers proposed that baseline models could provide risk stratification at the time of surgery, potentially informing decisions regarding adjuvant therapy. Dynamic models could subsequently update site-specific recurrence risk as new surveillance information became available.

For patients with colorectal cancer liver metastases undergoing curative-intent resection, this framework could support more individualised surveillance strategies by incorporating developments occurring during follow-up.

The study demonstrated the potential of machine learning to provide both an initial assessment of recurrence and overall survival and a mechanism for updating risk over time. The reported model performance indicated that incorporating longitudinal clinical information could improve prediction compared with baseline assessment alone.

Further application of this framework may help integrate surgical, pathological, molecular and surveillance data into personalised risk assessment for patients undergoing curative-intent hepatectomy for CRLM.

Reference

Winicki NM et al. Prediction of site-specific recurrence and overall survival for colorectal cancer with liver metastases. npj Precis Onc. 2026;DOI:10.1038/s41698-026-01716-3.