Depression diagnosis based on Deep Learning Using Time-series Sleep Quality Data.

Area of Science:

• Digital health
• Computational psychiatry
• Sleep science

Background:

• Depressive disorder is a prevalent global mental health condition with high chronicity and suicide risk.
• Current diagnosis relies on subjective methods like clinical interviews and questionnaires.
• There is a critical need for objective, digital biomarker-based diagnostic tools for depression.

Purpose of the Study:

• To develop and evaluate a deep learning-based multivariate time-series depression classification model.
• To utilize sleep data from wearable devices for objective depression diagnosis.
• To identify significant sleep biomarkers indicative of depressive disorder.

Main Methods:

• Employed deep learning architectures: MLSTM-FCN, InceptionTime, and Time-series Transformer.
• Extracted features from ten candidate sleep biomarkers derived from wearable device data.
• Classified depression status based on multivariate time-series sleep data.

Main Results:

• The MLSTM-FCN model achieved the highest performance with an Area Under the Curve (AUC) of 0.91.
• InceptionTime and Time-series Transformer models showed AUC scores of 0.82 and 0.78, respectively.
• Total time in bed, REM sleep latency, and light sleep duration were identified as significant depression indicators.

Conclusions:

• A deep learning model using wearable sleep data provides a cost-effective and objective method for depression diagnosis.
• The MLSTM-FCN model demonstrates high potential for clinical application in community settings.
• Objective sleep biomarker analysis can significantly aid in the diagnosis of depressive disorder.