A Machine Learning Model for Predicting Prognosis in HCC Patients With Diabetes After TACE.

Abstract

Purpose: Type II diabetes mellitus (T2DM) has been found to increase the mortality of patients with hepatocellular carcinoma (HCC). Therefore, this study aimed to establish and validate a machine learning-based explainable prediction model of prognosis in patients with HCC and T2DM undergoing transarterial chemoembolization (TACE).

Patients and methods: The prediction model was developed using data from the derivation cohort comprising patients from three medical centers, followed by external validation utilizing patient data extracted from another center. Further, five predictive models were employed to establish prognosis models for 1-, 2-, and 3-year survival, respectively. Prediction performance was assessed by the receiver operating characteristic (ROC), calibration, and decision curve analysis curves. Lastly, the SHapley Additive exPlanations (SHAP) method was used to interpret the final ML model.

Results: A total of 636 patients were included. Thirteen variables were selected for the model development. The final random survival forest (RSF) model exhibited excellent performance in the internal validation cohort, with areas under the ROC curve (AUCs) of 0.824, 0.853, and 0.810 in the 1-, 2-, and 3-year survival groups, respectively. This model also demonstrated remarkable discrimination in the external validation cohort, achieving AUCs of 0.862, 0.815, and 0.798 in the 1-, 2-, and 3-year survival groups, respectively. SHAP summary plots were also created to interpret the RSF model.

Conclusion: An RSF model with good predictive performance was developed by incorporating simple parameters. This prognostic prediction model may assist physicians in early clinical intervention and improve prognosis outcomes in patients with HCC and comorbid T2DM after TACE.