Multi-class classification model for psychiatric disorder discrimination

Area of Science:

• Neuroscience
• Psychiatry
• Computational Biology

Background:

• Psychiatric diagnoses rely on symptom-based approaches using tools like DSM and ICD.
• The search for objective biomarkers, including in psychiatry, is an ongoing area of research.

Purpose of the Study:

• To investigate if electroencephalogram (EEG) data can function as a biomarker for psychiatric diseases.
• To differentiate and classify psychiatric conditions using machine learning analysis of EEG data.

Main Methods:

• Analyzed EEG absolute power values from 19 electrodes across 4 frequency bands (alpha, beta, delta, theta).
• Utilized machine learning models including C5.0, random forest (RF), support vector machine (SVM), and artificial neural networks (ANN).
• Employed 5-fold cross-validation with 3 repetitions for model training and hyperparameter optimization on 80% of the data, with testing on the remaining 20%.

Main Results:

• Machine learning models achieved high classification accuracy between disease groups (e.g., C5.0 and SVM-radial at 0.841, RF at 0.762).
• Specific conditions such as ADHD, depression, and schizophrenia were notably well-differentiated with perfect F-scores and balanced accuracy.
• EEG data demonstrated feasibility as a biomarker for predicting the presence of psychiatric diseases.

Conclusions:

• EEG data holds significant potential as a biomarker for the detection and diagnosis of psychiatric disorders.
• Machine learning analysis of EEG provides a promising avenue for objective assessment in psychiatric evaluations.
• The study contributes to advancing machine learning applications in psychiatric research, particularly for differentiating specific conditions.