Comparative evaluation of regression and machine-learning models for hepatocellular carcinoma risk stratification across diverse aetiologies

ABSTRACTBackground and Aims We aimed to develop machine learning (ML) models for hepatocellular carcinoma (HCC) risk stratification in patients with cirrhosis and tested their ability to identify those with an annual incidence>3% justifying more intensive surveillance. Methods Data from three prospective cohorts (ANRS CO12 CirVir, CO22 Hepather, APHP CIRRAL) were used. All patients underwent semiannual ultrasound surveillance and were split into training and validation sets. HCC incidence was analyzed using a competing risk framework. A Single Tree (ST) model was developed using conditional decision trees, Random Forest (RF) models were built by aggregating 1,000 trees. A deep neural network (DNN)-based survival model was also applied. ML model performance was compared to established regression-based scores aMAP (age–male–ALBI–platelets) and FASTRAK (FAST-MRI for HCC suRveillance in pAtients with high risK of liver cancer) Results Among 4,867 patients with non-viral cirrhosis or resolved/controlled viral cirrhosis, 294 (9.2%) developed HCC over a median follow-up of 61.5 months (annual incidence: 1.99%). The ST model identified four key predictors, generating five distinct risk groups. These included patients with mildly impaired liver function or those with elevated GGT and low platelet counts. The RF and DNN approaches confirmed ST findings and delineated complex interactions among predictors. Performance metrics (C-index, Brier score, decision curve analysis) showed no significant advantage of ML models over aMAP and FASTRAK. Calibration was consistent across models. ML models identified higher proportions of patients with an annual HCC incidence>3% (ST 44%; DNN 37%; RF 30% vs. aMAP 36%; FASTRAK 29%) Conclusions ML-based algorithms did not outperform traditional scores but provided novel insights into variable interactions and helped identify clinically relevant patient subgroups with differing HCC risk profiles. Impact and implications Accurate stratification of hepatocellular carcinoma risk in cirrhosis is essential to optimize surveillance strategies, and this study provides a scientific rationale for exploring machine learning approaches to capture complex, non-linear interactions among clinical variables beyond traditional regression models. Although machine learning did not improve predictive performance over established scores, it revealed clinically meaningful risk subgroups defined by liver function, platelet count, and GGT, underscoring its value as an interpretative and hypothesis-generating tool. These results are particularly relevant for hepatologists and clinical researchers seeking to refine risk-adapted surveillance and to inform the design of future models or trials.