Predictions of Early Hospitalization of Diabetes Patients Based on Deep Learning: A Review

Abstract

Unmanaged diabetes can result in a number of complications that need to be hospitalised. Diabetes is a chronic disorder. With preventive treatment, outcomes may be improved through early prediction of diabetes-related hospitalisation using data-driven algorithms. Here, we examine recent advances in deep learning methods for anticipating readmissions and unexpected hospital stays in adult patients with diabetes. Firstly, we present an overview of the main factors that indicate the need for hospitalisation due to diabetic complications. The research on hospitalisation risk prediction using structured health data, such as demographics, prescriptions, test results, etc., using conventional machine learning techniques is then summarised. Using data from insurance claims and electronic health records, we then examine current research that has used deep learning models. It is emphasised that longitudinal data can be included using recurrent neural networks. Model architectures, training methods, and important data modalities are covered. The assessment also addresses deployment difficulty and the model's performance assessment on real-world datasets. Ultimately, potential paths forward include hybrid models that integrate data diversity, explainable predictions, and clinical knowledge. In order to provide evidence-based predictions of the risk of hospitalisation and readmission for diabetes patients, we examine the potential and constraints of recently developed deep learning algorithms in this review.