Identifying Probable Neurological Disorders with Explainable Machine Learning Techniques

As the global burden of neurological disorders increases, the need for new tools to diagnose brain disease has become increasingly critical. Several machine learning models have been developed to classify various brain disorders; however, ensuring that these models can provide clinically useful information is as important as good prediction performance. In this study, we evaluated the effectiveness of various machine learning models to accurately identify individuals with potential cerebral pathology using neuropsychological data collected during routine clinical care. Three machine learning models-Random Forest, XGBoost, and Graph Neural Network-were trained and evaluated. The models' performances were compared to identify the most effective approach. The model with the best performance was then used to generate explainability plots, offering insights into the key features that contribute to predictions. Our work shows that Graph Neural Network best suited for routine clinical data, where missing and imbalanced data is commonplace due to the prioritization of patient needs over data completeness.