Machine Learning-Based Model Used for Predicting the Risk of Hepatocellular Carcinoma in Patients with Chronic Hepatitis B.

Object: Currently, predictive models that effectively stratify the risk levels for hepatocellular carcinoma (HCC) are insufficient. Our study aimed to assess the 10-year cumulative risk of HCC among patients suffering from chronic hepatitis B (CHB) by employing an artificial neural network (ANN).

Methods: This research involved 1717 patients admitted to Beijing Ditan Hospital of Capital Medical University and the People's Liberation Army Fifth Medical Center. The training group included 1309 individuals from Beijing Ditan Hospital of Capital Medical University, whereas the validation group contained 408 individuals from the People's Liberation Army Fifth Medical Center. By performing a univariate analysis, we pinpointed factors that had an independent impact on the development of HCC, which were subsequently employed to create the ANN model. To evaluate the ANN model, we analyzed its predictive accuracy, discriminative performance, and clinical net benefit through measures including the area under the receiver operating characteristic curve (AUC), concordance index (C-index), and calibration curves.

Results: The cumulative incidence rates of HCC over a decade were observed to be 3.59% in the training cohort and 4.41% in the validation cohort. We incorporated nine distinct independent risk factors into the ANN model's development. Notably, in the training group, the area under the receiver operating characteristic (AUROC) curve for the ANN model was reported as 0.929 (95% CI 0.910-0.948), and the C-index was 0.917 (95% CI 0.907-0.927). These results were significantly superior to those of the mREACHE-B(0.700, 95% CI 0.639-0.761), mPAGE-B(0.800, 95% CI 0.757-0.844), HCC-RESCUE(0.787, 95% CI 0.732-0.837), CAMD(0.760, 95% CI 0.708-0.812), REAL-B(0.767, 95% CI 0.719-0.816), and PAGE-B(0.760, 95% CI 0.712-0.808) models (p < 0.001). The ANN model proficiently categorized patients into low-risk and high-risk groups based on their 10-year projections. In the training cohort, the positive predictive value (PPV) for the incidence of liver cancer in low-risk individuals was 92.5% (95% CI 0.921-0.939), whereas the negative predictive value (NPV) stood at 88.2% (95% CI 0.870-0.894). Among high-risk patients, the PPV reached 94.6% (95% CI 0.936-0.956) and the NPV was 90.2% (95% CI 0.897-0.917). These results were also confirmed in the independent validation cohort.

Conclusion: The model utilizing artificial neural networks demonstrates strong performance in personalized predictions and could assist in assessing the likelihood of a 10-year risk of HCC in patients suffering from CHB.