Hybrid Fully Connected Neural-Bidirectional Long Short-Term Memory Networks for Diabetic Complication Risk Prediction
Abstract
Accurate estimation of the risk level of chronic diseases such as heart disease, kidney disease, and retinopathy is necessary for early detection and appropriate treatment planning for diabetic patients. The traditional machine learning models, i.e., Bidirectional long short-term memory (BiLSTM) and fully connected neural networks (FCNN), have been extensively used for disease prediction but are typically burdened with high computational complexity and redundant feature dependencies. In this research work, a new prediction model is proposed based on structured text data from CSV files to determine the levels of disease risk. The proposed method outperforms BiLSTM and FCNN in classification performance, with improved performance metrics. In addition, we also perform feature reduction using random forest (RF) and explainable artificial intelligence (XAI) techniques, such as SHAP (SHapley Additive exPlanations), with the objective of obtaining the most informative features. Regardless of feature reduction, the proposed system still maintains the best performance, confirming its efficacy and resilience in risk prediction. The outcomes reveal the potential for combining advanced deep learning models with feature selection techniques to improve diabetic complication disease risk assessment and prediction.