A Stability Analysis of SHAP and LIME Under Feature Correlation for Migraine Prediction Using Deep Learning

Migraine is a prevalent neurological disorder affecting the life quality and daily functioning of people globally. Over the years, Machine Learning (ML) models have widely been used in migraine forecasting, but, due to their blackbox nature, clinicians treat them with distrust and do not consider them practical. To enhance the model's transparency, various explainable Artificial Intelligence (XAI) techniques like Shapley Additive explanations(SHAP) and Local Interpretable Model-agnostic Explanations(LIME) are increasingly being used to provide an explanation of the model decisions. However, correlated or redundant features badly affect the stability of SHAP or LIME explanations. In this study, we develop a deep learning-based prediction model for migraine prediction using structured clinical dataset information. To test the stability of explanations, synthetic features are incorporated to test stability to correlated features, and the results of SHAP and LIME explanations before and after correlation are discussed. To test the stability of explanations quantitatively for features, Spearman rank correlation is used. Experimental results show that SHAP has moderate stability with a correlation coefficient of 0.80 compared to LIME explanations, which show higher stability with a correlation coefficient of 0.93. These results indicate that the reliability of the performed explanations using the XAI approach appears to depend on the level of interpretability, namely global and local levels.