Abstract
BACKGROUND: Individualized risk stratification in hepatitis B virus-related acute-on-chronic liver failure (HBV-ACLF) remains challenging. This study aimed to develop and validate a multi-task deep learning model using longitudinal liver ultrasound images for prognosis prediction. METHODS: A total of 372 HBV-ACLF patients were retrospectively enrolled, and baseline (T0) and 5 days post-admission (T1) liver ultrasound images, clinical data, and 30-day outcome (survival/mortality) were collected. A Siamese U-net deep learning model (Siamese U-Net) was trained to automatically segment the liver region and predict 30-day mortality using longitudinal liver ultrasound images from the training cohort (n = 290). The model output and clinical predictors were integrated into a combined model via Cox regression, with a clinical model developed for comparison. Model performance was evaluated for segmentation and prediction in the validation cohort (n = 82). RESULTS: Siamese U-Net-generated masks achieve Dice Coefficients of 0.912 and 0.913 against manual delineation for T0 and T1 images in the validation cohort. The Siamese U-Net significantly outperformed the clinical model (p < 0.01), achieving a C-index of 0.795 and an AUC of 0.851 in the validation cohort. Calibration curves and decision curve analyses showed superior calibration and clinical utility. The combined model achieved a C-index of 0.834 and an AUC of 0.892, marginally improving the Siamese U-Net (p > 0.05) but significantly enhancing the clinical model (p < 0.01) in the validation cohort. CONCLUSIONS: The Siamese U-Net emerges as a promising tool in predicting prognosis for HBV-ACLF, thereby enhancing clinical decision-making and improving patient outcomes.