Machine learning analysis for recurrence prediction of colorectal cancer liver metastases

Recurrence of Colorectal Cancer Liver Metastases (CRLM) remains a major challenge impacting therapeutic outcomes and patient prognosis. This study aimed to evaluate the predictive value of liver metastases characteristics and other indicators for CRLM recurrence using machine learning models.

A total of 524 CRLM patients who underwent radical resection were stratified into recurrence and non-recurrence groups. Ten machine learning (ML) algorithms were employed to assess comprehensive predictive value of key variables selected via LASSO regression. SHapley Additive exPlanations (SHAP) were used for model interpretation and visualization. Additionally, a nomogram prediction model was developed based on the key variables selected.

Factors including liver metastases characteristics were associated with CRLM recurrence risk. Machine learning algorithms demonstrated robust predictive performs. Among the ten models, the Logistic Regression (LR) exhibited the best discriminative ability (AUC = 0.762, F1_Score = 0.72, Accuracy = 0.686, MCC = 0.364). The constructed nomogram model provided moderate recurrence discrimination (1 year: AUC (95 %CI) = 0.692 (0.642-0.742); 3 years: AUC (95 %CI) = 0.732 (0.681-0.782); 5 years: AUC (95 %CI) = 0.771 (0.705-0.838)). We further developed machine learning models specifically for liver-only recurrence of CRLM, among which Naive Bayes (NB) showed the best performance (AUC = 0.767), and constructed an effective prediction nomogram.

This study identified risk factors for overall recurrence and liver-only recurrence after radical resection for CRLM based on liver metastases characteristics and clinical-pathological factors, establishing predictive models that potentially aid in developing individualized treatment strategies.