Enhancing Schizophrenia Prediction Using Class Balancing and SHAP Explainability Techniques on EEG Data

Machine learning (ML) makes predictions or supports decision making based on data, achieving high accuracy, saving time and resources, and even running real-time analysis. However, one drawback of these models is the lack of transparency in complex models, reducing confidence in sensitive fields such as health. This paper analyzes electroencephalogram (EEG) data to predict schizophrenia in patients. Three classifiers are compared, considering Support Vector Machines(SVM), Adaptive Boosting (AdaBoost), and Extreme Gradient Boosting (XGBoost). Three metrics are used to measure the classification process, including accuracy (ACC), area under the curve (AUC) and F-1 score (F1). XAI is incorporated into the pipeline to identify relevant features. XGBoost is the model with the best performance in predicting schizophrenia cases, reaching an ACC = 0.93, AUC = 0.93 and F1 score = 0.92, outperforming the SVM and AdaBoost algorithms. The SHAP explainability technique was applied on the XGBoost model, identifying the sex, IQ, delta T6, and delta Pz waves as the most relevant characteristics in the prediction processes. Based on the data analysis, we found that schizophrenia causes an alteration of the delta wave in an EEG, which is different to other mental illnesses. On the other hand, the study generates a specific impact by showing that XGBoost presents better results in 3 validation metrics (ACC, AUC, and F1 Score) compared to SVM and AdaBoost. It is shown that balancing classes is essential to obtain better ML predictions.