A Comprehensive Investigation into Detecting Schizophrenia from EEG Signals Using a Machine Learning Approach

Abstract

Schizophrenia is a prevalent psychiatric condition that places significant clinical demands on both patients and their caregivers. An accurate and expeditious diagnosis is essential for the effective treatment of schizophrenia. In this regard, the identification of classification biomarkers has the potential to enhance comprehension of the neural underpinnings of schizophrenia and supplement clinical assessments. Recent years have seen an increase in research into the diagnostic and prognostic utility of Machine Learning (ML) techniques in schizophrenia. Several such studies have attempted to classify individuals with schizophrenia from healthy controls using neuroanatomical features. However, the range of neuroanatomical measures utilized in these investigations has been limited thus far. The objective of this study was to detect schizophrenia at an early stage using the largest EEG signal dataset to date, consisting of 193 patients. To compile this dataset, the three largest open-source EEG signal datasets were merged and processed. For the most accurate detection of schizophrenia from EEG signals, an ML array was utilized. With the Gradient Boosting Classifier (GBC) method, feature engineering, and model tuning, this research achieved one of the highest classification accuracies to date, 93.3%, among the other supervised ML models used in the study. In addition, the study’s results demonstrated that precision, recall, and f1 score were, respectively, 84.6%, 80%, and 82%. The obtained results from this thesis surpasses all previous works using EEG signal in terms of accuracy and number of subjects considered and the results were obtained only using supervised model which is computationally lighter than typical signal analyzing Convolutional Neural Network (CNN) models. This thesis concentrates on the robustness and significance of larger EEG signal datasets, as contemporary studies have implemented prediction strategies on relatively smaller datasets