BiTabNet_CFs: A Novel Hybrid Deep Learning Framework for Diabetes Mellitus Detection

The detection of diabetes mellitus (DM) involves identifying the presence of the body's inability to produce insulin which results in high blood sugar levels. The model faces challenges with ambiguous data inputs, ambiguous patient data, computational overhead, and reliance on high quality sensor data. The study presented a new pipeline for reliable and interpretable classification of DM with Bidirectional TabNet_CounterFactuals (BiTabNet_CFs) to overcome the existing limitations. The input data undergoes preprocessing, which includes quantile normalization and adaptive Box-Cox transformation. Feature selection is performed using the Boruta algorithm, which is a robust random forest-based method that discerns relevant features. To address class imbalance, CTGAN uses conditional vectors and log-frequency sampling to generate realistic synthetic samples, ensuring representation of insufficient classes. A hybrid Deep Learning (DL) model that combines TabNet and BiLSTM is used for classification. TabNet uses sparse attention to identify and transform relevant features while BiLSTM captures bidirectional dependencies to improve patterns recognition. It includes CounterFactuals (CFs) for model interpretability to simulate minimal changes to understand small interrupts on predictions. The proposed method enhances predictive accuracy, making it highly applicable for real world clinical decision support in diabetes management. The proposed BiTabNet_CFs model shows superior performance in DM classification by achieving an accuracy of 98.7%, a precision of 96.4%, a recall of 90.1%, and an F1-score of 89.2% which significantly overcomes the performance of existing methods.