On Applying GraphLime for Trustworthy Diabet Prediction

Flavia Costi, Darian M. Onchis, Codruta Istin, Adina Magda Florea, Eduard Hogea

Abstract: This paper aims to provide a comprehensive analysis of the benefits of employing GraphLIME (Local Interpretable Model Explanations for Graph Neural Networks) for reliable diabetes mellitus prediction. Our focus is on highlighting the advantages of integrating GraphLIME with a features attention-mechanism, compared to the standard pairing of deep learning neural networks with the original LIME explainability method. This system enabled us to develop an effective approach for identifying the most relevant features and applying the attention mechanism solely to those features. We conducted a detailed comparison of the performance metrics between the two approaches. By incorporating the attention mechanism, the model reached an accuracy of 92.6% in addressing the problem. The model's performance is thoroughly illustrated, with results further assessed using the Receiver Operating Characteristic (ROC) curve. Applying this technique to a dataset of 768 patients with or without diabetes mellitus, we enhanced the model's performance by over 18%.