Collaborative deep neural network for survival prediction of hepatitis patients using electronic health records

Hepatitis infection presents a global health threat with significant morbidity and mortality implications. Early and precise survival prediction plays an essential role in guiding clinical decisions and enhancing patient outcomes. Electronic health records (EHRs) contain extensive information regarding the medical history of patients. The increasing volume of data in EHRs means they are becoming more valuable for data-driven healthcare research. With the rapid growth of EHR data, sophisticated analytical techniques are progressively required to capture meaningful patterns and support predictive modeling. Recent advances in deep learning (DL) techniques have demonstrated the capability to learn feature representations from data and enhance model performance across diverse fields. In healthcare, DL techniques have also been successfully implemented for clinical event prediction, disease classification, and EHR-based analysis, accomplishing extraordinary results. These developments emphasize the ability of DL systems to enhance survival prediction in hepatitis patients. Therefore, this study presents an Ensemble Deep Learning Framework for Survival Prediction of Hepatitis Patients (EDLF-SPHP). The proposed model aims to accurately predict survival outcomes in hepatitis patients by leveraging EHR data. To achieve this, the proposed framework first performs preprocessing, where the input EHR data is prepared for analysis through data normalization, missing data handling, and target value encoding, which ensures consistency of the data for further analysis. For robust feature selection, the proposed EDLF-SPHP model incorporates spearman rank correlation and ANOVA methods to recognize and retain the most significant attributes. Finally, the soft voting-based ensemble classification is designed to ensure reliable survival prediction, integrating three complementary architectures: a sparse autoencoder, a bidirectional temporal convolutional network, and a deep belief network. By combining their predictive outputs, the model enhances overall classification performance and achieves higher accuracy in forecasting survival performances for hepatitis patients. Additionally, SHAP is utilized to explain the model’s predictions by attributing the contribution of each feature to the final output, thereby enhancing interpretability and providing both global and local insight into feature importance. The experimental result analysis of the EDLF-SPHP approach is applied against a benchmark hepatitis dataset, and the comparative outcomes demonstrated improved solution over the existing systems with respect to different measures.