A Novel Machine-Learning Approach to Predict Recurrence After Resection of Colorectal Liver Metastases.

Abstract

Surgical resection of hepatic metastases remains the only potentially curative treatment option for patients with colorectal liver metastases (CRLM). Widely adopted prognostic tools may oversimplify the impact of model parameters relative to long-term outcomes. Patients with CRLM who underwent a hepatectomy between 2001 and 2018 were identified in an international, multi-institutional database. Bootstrap resampling methodology used in tandem with multivariable mixed-effects logistic regression analysis was applied to construct a prediction model that was validated and compared with scores proposed by Fong and Vauthey. Among 1406 patients who underwent hepatic resection of CRLM, 842 (59.9%) had recurrence. The full model (based on age, sex, primary tumor location, T stage, receipt of chemotherapy before hepatectomy, lymph node metastases, number of metastatic lesions in the liver, size of the largest hepatic metastases, carcinoembryonic antigen [CEA] level and KRAS status) had good discriminative ability to predict 1-year (area under the receiver operating curve [AUC], 0.693; 95% confidence interval [CI], 0.684-0.704), 3-year (AUC, 0.669; 95% CI, 0.661-0.677), and 5-year (AUC, 0.669; 95% CI, 0.661-0.679) risk of recurrence. Studies analyzing validation cohorts demonstrated similar model performance, with excellent model accuracy. In contrast, the AUCs for the Fong and Vauthey scores to predict 1-year recurrence were only 0.527 (95% CI, 0.514-0.538) and 0.525 (95% CI, 0.514-0.533), respectively. Similar trends were noted for 3- and 5-year recurrence. The proposed clinical score, derived via machine learning, which included clinical characteristics and morphologic data, as well as information on KRAS status, accurately predicted recurrence after CRLM resection with good discrimination and prognostic ability.