Prospective evaluation of speech as a digital biomarker for covert hepatic encephalopathy

Covert hepatic encephalopathy is a highly prevalent complication of liver failure or portosystemic shunting. However, the current diagnostic methods have not been widely adopted in routine clinical practice due to their time-consuming or technically complex nature. This study aimed to develop and evaluate a clinically feasible machine learning model capable of detecting covert hepatic encephalopathy using features extracted from sustained vowel phonation recordings. The XGBoost model achieved the highest AUROC of 81.20 (95% CI 73.03-89.73) and showed reasonable calibration across the full spectrum of predicted probabilities (Hosmer-Lemeshow p = 0.66). The most influential features indicated that patients with covert hepatic encephalopathy tended to have reduced variation in volume, maintaining more constant and harsher vocal quality.