An artificial neural networking model for the prediction of post-hepatectomy survival of patients with early hepatocellular carcinoma

Background and Aims: This study aimed to establish a prognostic artificial neural network model (ANN) for early hepatocellular carcinoma (HCC) following partial hepatectomy. Methods: Consecutive patients who were operated between February 2005 and March 2012 were prospectively studied. Seventy-five and 25% of these patients were randomly selected as a training cohort and an internal validation cohort. Similar patients from another hospital formed an external validation cohort. The predictive accuracy of the ANN for postoperative survival was measured by the area under the curve (AUC) on receiver operating characteristic (ROC) curve analysis. The results were compared with those obtained using the conventional Cox proportional hazard model, and the Hepato-Pancreato-Biliary Association (IHPBA), TNM 6th, and Barcelona-Clinic-Liver-Cancer (BCLC) staging systems. Results: The number of patients in the training, internal validation and external validation cohorts were 543, 182, and 104, respectively. On linear regression analysis, tumor size, number, alpha fetoprotein, microvascular invasion, and tumor capsule were independent factors affecting survival (P < 0.05). The ANN model was established based on these factors. In the training cohort, the AUC of the ANN was larger than that of the Cox model (0.855 vs 0.826, P = 0.0115), and the staging systems (0.784 vs TNM 6th: 0.639, BCLC: 0.612, IHPBA: 0.711, P < 0.0001 for all). These findings were confirmed with the internal and external validation cohorts. Conclusion: The ANN was significantly better than the other commonly used model and systems in predicting survival of patients with early HCC who underwent partial hepatectomy.