Machine Learning-Based Survival Analysis for Patients Receiving Lenvatinib for Unresectable Hepatocellular Carcinoma.

Purpose: Lenvatinib is an effective treatment for patients with intermediate- to advanced-stage unresectable hepatocellular carcinoma (HCC). However, tumor response and survival outcomes vary widely. Traditional machine learning (ML) models have been developed to predict treatment response or survival status at discrete time points. However, an overall prediction of overall survival (OS) and progression-free survival (PFS) incorporating censored survival data is lacking. We aimed to conduct a comprehensive survival analysis of OS and PFS by using ML-based survival models. Patients and Methods: This multicenter, retrospective study included patients with unresectable HCC receiving lenvatinib across five healthcare centers. Demographic data, laboratory results, tumor characteristics, and survival outcomes were collected. Five ML-based survival models were developed and compared using Harrell's concordance index (C-index). The predicted risk scores were used to stratify patients into low-, intermediate-, and high-risk groups and validated in the test set. Results: 205 patients were included for training and validation. Among the five ML models, the GBM-Cox model achieved the highest C-indices for both OS (0.617) and PFS (0.645) prediction. The predicted risk scores stratified the patients into low-, intermediate-, and high-risk groups for OS (median, 18.7 vs 13.6 vs 8.8 months; Conclusion: ML-based survival models successfully stratified patients into low-, intermediate-, and high-risk groups for OS and PFS. Key features included ALBI score and alanine aminotransferase for OS, and macrovascular invasion and ALBI score for PFS. These models have the potential to guide clinicians' treatment decisions and provide prognostic evaluations. Future prospective studies with larger cohorts, as well as integration of imaging biomarkers are warranted to optimize these predictive models.