Construction of a predictive model for the risk of moderate-to-severe cancer-related fatigue in colorectal cancer chemotherapy patients: an interpretable machine learning approach.

Purpose: This study aimed to analyze the influencing factors of moderate-to-severe cancer-related fatigue (CRF) in colorectal cancer (CRC) chemotherapy patients and to develop a predictive risk stratification model.

Methods: A total of 630 CRC chemotherapy patients were selected from five hospitals in China. Data were collected using a general information forms, the Piper Fatigue Scale-Revised (PFS-R), the Hospital Anxiety and Depression Scale (HADS), and the Pittsburgh Sleep Quality Index (PSQI). The data was randomly divided into a training set and a test set in a 7:3 ratio, and feature selection was performed using univariate analysis and LASSO regression. Five machine learning algorithms were used to construct moderate-to-severe CRF models. The Shapley additive explanation (SHAP) method is used to increase the interpretability of the optimal performance model.

Results: The overall incidence of moderate-to-severe CRF was 70.5%. The random forest (RF) model performed the best, with an AUC of 0.906, sensitivity of 0.943, accuracy of 0.931, precision of 0.977, specificity of 0.848, and F1 score of 0.960. Based on the analysis of the absolute mean SHAP values, the feature importance of the RF model, from highest to lowest, was sleep quality score, anxiety score, anorexia, magnesium ion concentration, smoking history, place of residence, and cancer stage.

Conclusions: The RF model demonstrated superior predictive performance, positioning it as a viable screening tool for assessing the risk of moderate-to-severe CRF in CRC patients receiving chemotherapy. This approach may facilitate early intervention and improve clinical management of CRF symptoms.