Development and validation of machine learning models for diagnosing hepatocellular carcinoma risk and survival in patients with diabetic cirrhosis.

Diabetes mellitus and liver cirrhosis exhibit a bidirectional interaction, which significantly elevates the risk of hepatocellular carcinoma (HCC) in this susceptible population and increases the difficulty of clinical diagnosis and treatment. This study established and validated reliable and practical models for HCC diagnosis and all-cause mortality prognosis. Meanwhile, we constructed prognostic models specific to HCC, liver-related death, and diabetes-related death, enabling a comprehensive analysis of independent risk factors for different outcomes. A total of 307 patients with diabetic cirrhosis were recruited from Jurong Hospital Affiliated to Jiangsu University, and their clinical data were collected for the study. Eight machine learning algorithms were tested, among which the Gradient Boosting Decision Tree model showed the optimal diagnostic performance with an area under the receiver operating characteristic curve (AUC) of 0.968. A six-variable nomogram incorporating sex, alkaline phosphatase, serum sodium, apolipoprotein A1, direct bilirubin, and cholinesterase (CHE) was developed. The nomogram yielded a mean AUC of 0.724 via 1000 bootstrap resampling validations, an AUC of 0.779 in internal validation, and an AUC of 0.814 in external validation. Subgroup validation confirmed the nomogram’s stable efficacy in the hepatitis B virus (HBV)-infected subgroup (AUC = 0.704), hepatitis C virus (HCV)-infected subgroup (AUC = 0.826), and non-viral aetiology subgroup (AUC = 0.778). Through univariate Cox, Least Absolute Shrinkage and Selection Operator (LASSO), and multivariate Cox regression analyses, HCC, age, CHE, and lactate dehydrogenase (LDH) were identified as independent prognostic factors for all-cause mortality. A prognostic nomogram incorporating these variables along with sex exhibited good stability and extrapolability, with 1-, 3-, and 5-year AUCs of 0.848, 0.833, and 0.841 in the internal validation cohort, and 0.796, 0.781, and 0.759 in the external validation cohort, respectively. Three disease-specific mortality models were subsequently developed. A Fine-Gray model for HCC-specific mortality identified CHE (P = 0.018) and creatinine (P = 0.038) as independent predictors. Cumulative Incidence Function (CIF) analysis showed that during the 5000-day follow-up period, the cumulative incidence of non-HCC death rapidly rose to nearly 47%, while that of HCC death remained relatively stable at approximately 13%, indicating that the risk of non-HCC death was much higher than that of HCC death. Furthermore, a cause-specific Cox model for liver-related mortality (CS-Cox (Liver)) confirmed age, LDH (P = 0.012), and blood glucose (GLU) (P = 0.013) as key influencing factors. Similarly, the cause-specific model for diabetes-related mortality (CS-Cox (DM)) demonstrated that age (P < 0.001) and neutrophil-to-hemoglobin ratio (NHR) (P = 0.021) were independent prognostic variables. Decision Curve Analysis (DCA) verified that both the CS-Cox (Liver) and CS-Cox (DM) models had favorable clinical net benefits across most threshold probabilities, supporting their potential utility in clinical practice. This study provides important insights for HCC diagnosis and survival analysis in this high-risk population, and emphasizes the necessity of balancing HCC surveillance and comorbidity management in clinical practice.