The team of Dr hab. Jan Kamiński (Laboratory of Neurophysiology of Mind) has just published a systematic review on the use of AI for EEG-based psychiatric diagnosis.
This is essential reading for anyone working at the intersection of neuroscience and machine learning.
For those looking to enter this research area, the review offers several key takeaways:
- Sufficient sample size and subject-wise validation are crucial to avoid overfitting.
- Line-noise filtering during preprocessing may improve classification accuracy, unlike other artifact-reduction methods, which showed no significant effect.
- Combining feature types is beneficial — integrating frequency- and time-domain EEG characteristics appears especially advantageous.
- Feature selection helps mainly with complex feature sets and traditional ML models, but seems less necessary for neural networks.
- Neural networks generally outperform traditional ML approaches, with attention-based networks and CNN-LSTM models emerging as the most effective architectures.
- The most discriminative EEG features include connectivity measures, frontal-region activity, and alpha/beta bands for depression; raw EEG or connectivity features and theta/alpha bands for schizophrenia; and theta/alpha bands for addiction.
- Classification of PTSD from resting-state EEG appears to be the least effective among the disorders reviewed.
Well worth a read for anyone interested in the future of AI-assisted psychiatric diagnostics.