Hepatic encephalopathy (HE) is an independent risk factor for mortality in patients with hepatitis B virus-related acute-on-chronic liver failure (HBV-ACLF). We aim to assess the effectiveness of 3D deep learning based on CT imaging in predicting HE in HBV-ACLF patients. A retrospective study was conducted on 222 HBV-ACLF patients who underwent CT scans at two medical centers. The cohort from Huashan Hospital, consisting of 146 patients, was randomly divided into a training cohort (n = 102) and a validation cohort (n = 44). An external test cohort (n = 76) was sourced from Shanghai Public Health Clinical Center. We employed two architectures, ResNet50 and DenseNet201, to extract 3D deep learning features related to HE. Classical radiomics features were extracted from liver CT images and modeled using three machine learning algorithms. Clinical features associated with HE were analyzed via multivariate logistic regression. The 3D deep learning model demonstrated strong performance in predicting HE in ACLF patients, particularly the ResNet50 model, which achieved AUC values of 0.844 (95% CI: 0.752-0.936) in the training cohort, 0.833 (95% CI: 0.671-0.996) in the validation cohort, and 0.833 (95% CI: 0.686-0.981) in the test cohort. Classical radiomics also demonstrated predictive potential, where XGBoost outperformed other algorithms. INR and ammonia levels were independently associated with HE. The nomogram model integrating 3D deep learning, radiomics features, and clinical predictors achieved the highest predictive accuracy across all cohorts. 3D deep learning performed excellently in predicting HE in ACLF patients, and its combination with radiomics and clinical features further enhanced predictive performance.