Artificial Intelligence in Psychiatry: ML and DL Models for Schizophrenia (SCZ) Detection Using EEG Signals

Abstract

Schizophrenia (SCZ) affects about 1 % of the global population and can manifest as chaotic thoughts, vivid hallucinations, and firmly held false beliefs. Because early, accurate diagnosis dramatically improves treatment outcomes, researchers are turning to artificial-intelligence methods that can read brain-wave recordings and flag the disorder automatically. In this study we evaluated a range of machine-learning (ML) and deep-learning (DL) approaches on electroencephalogram (EEG) data collected from 150 patients with schizophrenia and 150 healthy control participants. We extracted three types of information from each recording: Time-domain metrics – simple statistics such as mean, variance, and signal-shape features. Frequency-domain characteristics – power in standard EEG bands (delta, theta, alpha, beta, gamma). Time-frequency representations – spectrograms that capture how the frequency content evolves over time. We then trained several classifiers, from classic algorithms like Support-Vector Machines (SVM) and Random Forests to modern neural networks, including stand-alone Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) recurrent nets, and a hybrid CNN-LSTM model that combines spatial feature extraction with temporal sequence learning. The results were clear: deep-learning models, especially the CNN-LSTM hybrid, outperformed the traditional methods. The best model achieved more than 94 % overall accuracy, with a sensitivity of 93.2 % (correctly identifying patients) and a specificity of 95.1 % (correctly rejecting healthy subjects). These findings reinforce the promise of AI-driven diagnostics in psychiatry, suggesting that sophisticated EEG-based tools could soon become valuable companions to clinicians, helping to diagnose schizophrenia faster and more reliably.