Construction of a Bayesian network-based risk prediction model for hepatocellular carcinoma in cirrhotic patients

To investigate clinical data of hospitalised cirrhosis patients, identify risk factors for cirrhosis progression to hepatocellular carcinoma, establish a risk prediction model, and provide scientific basis for early identification of high-risk patients.

Hospitalised cirrhosis patients treated at Xinjiang Uygur Autonomous Region People's Hospital between January 2019 and December 2023 were selected. Their inpatient examination records were retrieved, including medical record summaries alongside results for coagulation function, complete blood count, liver function tests, urinalysis, renal function tests, tumour markers, comprehensive hepatitis panel, comprehensive thyroid function panel, comprehensive biochemical panel, glucose series, and lipid series. Patients diagnosed with hepatic malignancy during subsequent hospitalisations (excluding the initial admission) formed the cancer progression group, while those without hepatic malignancy constituted the control group. Univariate and multivariate analyses identified risk factors for hepatocellular carcinoma (HCC) progression in cirrhosis patients. A predictive model for HCC development in cirrhosis patients was constructed using a combined Lasso regression model and Bayesian network model.

This study enrolled 1,204 individuals, including 1,128 cirrhosis patients, of whom 76 progressed to liver malignancy. Multivariate logistic regression analysis indicated that female gender was a protective factor against cirrhosis progression to liver malignancy(OR = 0.532, OR95% CI = 0.297-0.952); while hepatitis B, elevated total cholesterol, and reduced antithrombin III activity were risk factors for progression to hepatocellular carcinoma (OR = 4.080, OR95%CI = 2.443-6.814;OR = 2.308, OR95%CI = 1.132-4.707, OR = 2.982, OR 95% CI = 1.389-6.402) (P < 0.05); The LASSO regression model ultimately identified 14 variables most significantly associated with the transformation of cirrhosis into malignant liver tumours: PT, TT, MAO, ALP, PDW, CRP, CK-MB, Ca, P, TC, AFP, FT4, SCC, AT3. The results of the Bayesian network model indicate that DOI, AT3, SCC, CRP, and TC have a direct connection with the occurrence of hepatic malignancy; Gender, FT4, hepatitis type, and Ca indirectly influence the development of hepatic malignancy. Using sensitivity analysis, we examined the probability of cirrhosis patients progressing to hepatocellular carcinoma under specific conditions. The results indicated that the probability of liver cirrhosis patients progressing to hepatocellular carcinoma was 0.0749 (7.5%) when AT3 levels were normal, 0.0709 (7.1%) when AT3 was low, and 0.0850 (8.5%) when AT3 was high. When TC levels were normal, the probability was 0.0763 (7.6%), and it increased to 0.1309 (13.1%) when TC was elevated. The performance of the Bayesian network model (AUC = 0.857, Brier Score = 0.052, Accuracy = 0.940, C-index = 0.833 (95% CI: 0.784-0.882), Sensitivity = 0.836, Specificity=0.744) was superior to that of the logistic regression model (AUC = 0.780, Brier Score = 0.056, Accuracy = 0.934, C-index = 0.785 (95% CI: 0.726-0.843), Sensitivity = 0.727, Specificity = 0.766). Ten-fold cross-validation showed an average accuracy of 0.93. After balancing sensitivity and specificity, the optimal threshold was determined by maximizing the Youden index (0.052), with a predicted probability >0.052 indicating progression from liver cirrhosis to hepatic malignancy. Based on the ROC curve, threshold 1 was set at 0.2 and threshold 2 at 0.8, establishing risk stratification as follows: low risk (predicted probability <0.2), intermediate risk (0.2 ≤ predicted probability ≤ 0.8), and high risk (predicted probability >0.8). This resulted in 318 patients classified as low risk, 42 as intermediate risk, and 1 as high risk.

Gender, hepatitis B, TC, and AT3 constitute risk factors for hepatocellular carcinoma in cirrhotic patients; Gender, hepatitis type, DOI, FT4, AT3, SCC, CRP, MAO, and Ca are associated with the progression of liver cirrhosis to malignant liver tumours either directly or indirectly. The risk prediction model constructed by combining LASSO regression with Bayesian networks demonstrates good predictive value.