Towards Improved Machine Learning Models for Adult-onset Psychiatric Disorders Classification using resting-state EEG: A Systematic Review

Recent advances in machine learning (ML) have led to growing interest in the classification of psychiatric disorders using biomarkers such as electroencephalography (EEG), offering promising prospects for improving diagnostic accuracy. However, the methodological variability in existing studies makes it challenging to identify the best strategies for enhancing this crucial tool. This systematic review summarizes studies using resting-state EEG for psychiatric disorders classification to identify key strategies for optimizing classification performance. Our analysis reveals a significant interaction between the EEG features used for classification and the chosen ML algorithms. Notably, neural networks outperform traditional ML methods, especially when applied to raw data or to complex data without feature selection. Relying solely on linear features can undermine model performance, whereas combining diverse feature types leads to higher accuracies. Additionally, preprocessing data with a notch filter can enhance model performance. We underline the importance of obtaining sufficient sample sizes and using subject-wise validation to mitigate potential overfitting. Moreover, we investigate the best-performing features, showing the importance of connectivity features, alpha and beta frequency bands, and frontal brain regions for depression detection; raw features and theta and alpha frequency bands for schizophrenia; and combined features and theta and alpha frequency bands for addiction. These insights synthesize the most effective approaches and provide valuable guidance for developing new tools in this field.