COMPARISON OF DEEP LEARNING MODELS LSTM AND BILSTM IN DIABETES PREDICTION

Abstract

Diabetes mellitus remains a major global health concern, requiring early detection to prevent severe complications and reduce mortality. This study developed and evaluated two deep learning architectures, Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM), for diabetes prediction using the Pima Indians Diabetes Dataset. The research methodology involved systematic preprocessing, including outlier handling with median imputation, data normalization, and training–testing data splitting (80:20). Both models were trained using 614 samples for training and 154 samples for testing, with 50 epochs and a batch size of 32. The evaluation was performed using accuracy, precision, recall, F1-score, and AUC metrics. Results indicated that LSTM achieved an accuracy of 74.03%, while BiLSTM slightly outperformed it with 74.68%. Confusion matrix analysis further revealed that BiLSTM reduced false negatives and provided more consistent learning stability compared to LSTM. Accuracy and loss curves confirmed BiLSTM’s superior convergence and generalization capability. These findings demonstrate that BiLSTM is more effective and reliable for diabetes prediction tasks. The study concludes that BiLSTM offers better potential for integration into decision-support systems, and future research could enhance performance through larger datasets, advanced optimization, and real-world clinical validation.