Abstract
BACKGROUND: Hepatic encephalopathy (HE) contributes significantly to mortality among patients with liver cirrhosis. Early prediction of HE is essential for clinical decision-making, yet remains challenging-particularly in noncancer-related cirrhosis due to the unpredictable disease course. OBJECTIVE: This study aimed to develop a novel machine learning (ML) model to improve early prediction of HE in patients with noncancer-related cirrhosis. METHODS: A multicenter, retrospective cohort study was conducted from January 2010 to December 2017 across all Chang Gung Memorial Hospital branches in northern, middle, and southern Taiwan. We applied several ML models to evaluate HE predictability and compared their performance in the training dataset and testing dataset. Optimal sensitivity and specificity were determined using the Youden index. The best ML model was interpreted by the Shapley Additive Explanations plot. RESULTS: A total of 5878 patients with cirrhosis were included in the analysis, of whom 1187 (20.2%) subsequently developed HE. Compared to the non-HE group, patients with HE were older (median age 55, IQR 46-65 vs median age 54, IQR 44-66 years; P=.04) and had higher rates of hepatitis B virus infection (351/1187, 30% vs 961/4691, 20.5%; P<.001), alcohol use (540/1187, 45.5% vs 1512/4691, 32.2%; P<.001), sepsis (393/1187, 33.1% vs 792/4691, 16.9%; P<.001), and mortality (425/1187, 35.8% vs 502/4691, 10.7%; P<.001), along with distinct laboratory abnormalities reflecting liver dysfunction. Among the ML algorithms evaluated, the extreme gradient boosting algorithm demonstrated the highest predictive accuracy, achieving an area under the curve (AUC) of 0.86 (95% CI 0.83-0.88) in the testing dataset. This performance was significantly superior to that of the neural network (AUC 0.79, 95% CI 0.76-0.81; P<.001), support vector machine (AUC 0.77, 95% CI 0.73-0.80; P<.001), and the model for end-stage liver disease score (AUC 0.74, 95% CI 0.71-0.77; P<.001). Using a probability threshold of 0.25, the extreme gradient boosting model demonstrated a sensitivity of 72% (95% CI 0.67-0.77), specificity of 80% (95% CI 0.78-0.82), a positive predictive value of 48% (95% CI 43-53), and a negative predictive value of 92% (95% CI 90-94) in the testing set. Comparable performance was observed in the training dataset, with a sensitivity of 80% (95% CI 0.77-0.83), specificity of 81% (95% CI 0.80-0.82), and a negative predictive value of 94% at the same threshold. The most influential predictive variables identified by the model included serum ammonia, aspartate transaminase, alanine transaminase, prothrombin time, and serum potassium. CONCLUSIONS: We developed a novel ML model for predicting HE in patients with noncancer-related cirrhosis. This model provides a practical guide to help physicians and these patients in shared decision-making regarding treatment strategy, with the ultimate goal of improving clinical care and reducing the burden of HE-related morbid complications.