Multimodal deep learning for detecting current depressive symptom status using wearable time-series data and...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Occupational Health

Background:

• Mental health in high-risk occupations is often underreported due to stigma.
• Wearable sensors capture dynamic data, while brain MRI offers stable biological insights.
• Integrating these modalities may improve objective mental health assessment.

Purpose of the Study:

• To evaluate a multimodal deep learning model combining wearable data and structural brain MRI radiomics.
• To identify current depressive symptom status in firefighters and prosecution investigators.
• To assess the discriminative performance of the multimodal approach compared to wearable-only models.

Main Methods:

• A total of 291 participants (firefighters and investigators) were monitored for ~4.8 weeks.
• Structural brain MRI radiomics and time-series wearable data were used.
• Temporal deep learning models, including LSTM, were trained and evaluated.

Main Results:

• The multimodal deep learning model achieved an AUROC of 0.867.
• Multimodal models significantly outperformed wearable-only models.
• Key contributors included specific brain regions (left paracentral lobule, amygdala) and sleep oxygen saturation.

Conclusions:

• A multimodal approach integrating wearable and brain MRI data can objectively identify depressive symptoms in high-risk occupations.
• This method offers a supportive tool for mental health screening in occupational settings.
• Neuroanatomical and behavioral data together enhance diagnostic accuracy.