Related Experiment Video
Updated: Aug 6, 2026
Closed-Loop Neurostimulation for Biomarker-Driven, Personalized Treatment of Major Depressive Disorder
Published on: July 7, 2023
Development of a Machine Learning-Based Model for Classifying Depression Using Physiological and Psychological
Sooah Jang1,2, Jinsoo Park1, HyunKyung Shin1
1Research Institute of Minds.AI, Co. Ltd., Seoul, Republic of Korea.
Objective:
This study aimed to develop and evaluate machine learning-based models for classifying depression symptoms using salivary hormone markers (cortisol and dehydroepiandrosterone [DHEA]) and psychological indicators from depression screening devices, and to determine the extent to which salivary hormone markers contribute to depression symptom classification.
Methods:
Two models were developed, a multi-class classification (four degrees of depression) and a binary classification model (normal vs. depression), using psychological indicators that assess depression symptoms, protective and vulnerability factors, and physiological indicators including salivary cortisol and DHEA. Each model was evaluated both with and without physiological indicators. Data from 368 individuals were used for model training, with 92 individuals reserved for independent testing. Out of 36 variables, including psychological indicators from the PROtective and Vulnerable factors battEry test and physiological markers (e.g., salivary cortisol and DHEA), 24 non-demographic variables with independent contributions were selected, excluding demographic information. To ensure model robustness, overfitting risks were mitigated using k-fold cross-validation, class weight balancing, and grid search optimization. Model performance was evaluated based on accuracy, precision, recall, and receiver operating characteristic values.
Results:
The multi-class classification model achieved 85.9% and 76.1% accuracy with and without physiological indicators, respectively. The binary classification model (depression presence or absence) achieved 97.8% accuracy both with and without physiological indicators.
Conclusion:
The machine learning models demonstrated high accuracy in depression classification, showing a noteworthy improvement in classification performance when integrating psychological and physiological markers. Future clinical trials will be conducted to secure additional clinical data to verify the model's reliability and clinical validity.
