Predicting alcohol use disorder risk in firefighters using a multimodal deep learning model: a cross-sectional study

Area of Science:

• Neuroimaging and computational psychiatry
• Occupational health and safety
• Deep learning applications in clinical diagnostics

Background:

• Firefighters face elevated alcohol use disorder (AUD) risk due to trauma exposure.
• Traditional self-report screening is hampered by stigma and career concerns, leading to underreporting.
• Objective AUD risk stratification is needed for this high-risk occupational group.

Purpose of the Study:

• To develop and validate a multimodal deep learning framework for objective AUD risk stratification in firefighters.
• To integrate T1-weighted structural MRI with neuropsychological assessments for enhanced accuracy.
• To avoid computationally intensive functional neuroimaging protocols.

Main Methods:

• A cross-sectional study analyzed 689 active-duty firefighters.
• Structural MRI and neuropsychological tests (Grooved Pegboard, Trail Making Test) were employed.
• A deep learning model combined ResNet-50, Vision Transformers, and clinical variables, with interpretability analysis.

Main Results:

• The multimodal framework achieved 79.88% accuracy and 79.65% AUC, significantly outperforming clinical-only and neuroimaging-only models.
• Cross-modal integration yielded a 17.35 percentage-point improvement in performance.
• Interpretability highlighted the importance of biological sex and motor coordination metrics.

Conclusions:

• Structural neuroimaging and neuropsychological assessment offer a pragmatic, objective approach to AUD screening in high-risk professions.
• This method reduces acquisition time and computational needs compared to complex protocols.
• The framework has broader implications for psychiatric risk stratification in trauma-exposed populations.