FedBiCNN: Federated BiLSTM-CNN Framework for Early Detection of Type III Diabetes Mellitus

Abstract

Type III Diabetes Mellitus (T3DM) is characterized by insulin resistance in the brain and neurodegenerative disorders, and is the main objective of timely and accurate detection to ensure early clinical intervention. But machine learning systems with central control are not without troubles due to patient privacy, data diversity and difficulties in generalizing across distributed healthcare institutions. To solve these problems, this study suggests a Federated Learning Enabled BiLSTM-CNN Framework (FedBiCNN) for early diagnosis of Type III Diabetes Mellitus (T3DM) based on data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). The proposed architecture consists of data preprocessing, data normalization, a federated learning aggregation server using FedAvg algorithm, CNN local feature extraction, BiLSTM temporal pattern learning, an attention mechanism for clinical feature selection, explainable artificial intelligence using SHAP for model interpretability and a global model aggregation strategy for robust prediction. The framework was simulated and tested in the TensorFlow Federated (TFF) simulation environment. Experimental results show that FedBiCNN has average accuracy of 98.74%, MCC of 0.974, response time of 0.81 s, training time of 118.6 s and communication cost of 42.5 MB. Compared with CNN-LSTM, FedCNN, and ResNet-50, the proposed model improved accuracy by 5.72%, 4.19%, and 3.06%, respectively, while reducing response time by 38.18%, 29.57%, and 20.59%, training time by 31.04%, 23.14%, and 16.18%, and communication cost by 35.90%, 27.35%, and 18.11%, respectively. The results show that FedBiCNN is a privacy-preserving, interpretable, and highly efficient framework for early prediction of Type III diabetes that facilitates reliable clinical decision-making and enhances federated healthcare intelligence systems.