Related Experiment Videos
Multimodal machine learning model integrating electroencephalography, heart rate variability and clinical measures
Byeongjae Kang1, Ji-Yoon Lee2, Jeong Woo Chang3
1Department of Digital Health, SAIHST, Sungkyunkwan University, Seoul, Republic of Korea; Medical AI Research Center, Research Institute for Future Medicine, Samsung Medical Center, Seoul, Republic of Korea.
Background:
This study aimed to develop and evaluate multimodal machine learning (ML) approaches for differentiating bipolar disorder (BD), major depressive disorder (MDD), and schizophrenia (SZ), conditions that are clinically difficult to distinguish because of overlapping symptom presentations despite their distinct long-term treatment and management needs.
Methods:
We analyzed a retrospective cohort of 1035 patients (406 BD, 522 MDD, and 107 SZ) using multimodal data, including demographics, clinical scales, heart rate variability, and high-dimensional sensor- and source-level electroencephalography (EEG). To reduce the dimensionality of EEG features, we applied a two-step feature selection procedure combining analysis of variance and recursive feature elimination with cross-validation. We then trained and compared the diagnostic performance of unimodal versus multimodal ML models.
Results:
The multimodal ML model based on the CatBoost algorithm achieved an AUROC of 0.776, outperforming all unimodal approaches in differentiating these psychiatric disorders. Delta-, alpha-, and gamma-band phase lag and coherence between frontal, central, and occipital regions emerged as important EEG connectivity features for distinguishing the disorders.
Conclusions:
These findings suggest that multimodal ML may improve the differentiation of SZ, BD, and MDD and may have potential as a decision-support tool, although external and prospective validation is required before clinical implementation.