DHM-Net: Smart Healthcare-Enabled Diabetes Prediction Using Temporal Pattern Learning and Adaptive Data Balancing

Diabetes is a global health problem and a leading cause of early death due to its ability to cause severe complications and organ failure if undetected. Traditional diagnostic methods are based on manual assessment, which can lead to human error; hence, advanced computational methods are increasingly incorporated into clinical decision support systems. While deep learning (DL) models have been used in a wide range of healthcare applications, challenges remain, including learning intricate patterns, handling unbalanced datasets, and interpreting predictions. We propose a DHM-Net (Diabetes Health Monitoring Network) model that integrates Long Short-Term Memory (LSTM) and DenseNet169 for diabetes prediction. The model's ability to learn temporal patterns in the data and capture intricate relationships between features helps to increase the precision of diabetes prediction. An adaptive synthetic oversampling technique is applied to the Diabetes Health Indicators dataset to address class imbalance. Additionally, the robustness and generalization potential of the proposed method are considered using k-fold cross-validation. To improve data quality and prevent redundancy, duplicate entries are removed during the preprocessing stage. Perceptron-based learning is incorporated into the suggested framework to address issues with class imbalance, redundant data, and limited interpretability in early diabetes prediction. The experimental results show that the model outperforms several baseline and current deep learning models for diabetes prediction.