A deep learning model trained on brain scans and cognitive tests could help protect firefighters at risk of alcoholism, changing how high-stress professions approach mental health.
Study: Predicting alcohol use disorder risk in firefighters using a multimodal deep learning model: a cross-sectional study. Image credit: Tsuguliev/Shutterstock.com
Firefighters undergo chronic and repeated trauma in their professional lives, putting them at high risk for alcohol use disorder (AUD). A recent study published in Frontiers in Psychiatry examines the use of a deep learning multimodal framework to assess AUD risk objectively.
Stress fuels addiction
Firefighters face continuous exposure to emergencies and disasters, leading to significant and cumulative mental stress. The intense physical demands and psychological strain of their work make firefighters particularly vulnerable to developing mental health disorders, especially AUD.
Epidemiologic evidence shows a much higher rate of mental health screening among public safety staff compared to the general population. One in seven people screens positive for one or more mental disorders, while ~27 % have two or more.
Alcohol is often used as a maladaptive coping mechanism to help deal with trauma and stress. Firefighters may turn to alcohol for temporary relief, as it can reduce hyperarousal, dull intrusive memories of traumatic events, and ease emotional distress. Firefighter culture normalizes drinking when coming off active duty, but also discourages displays of vulnerability. This leads to a situation in which it is both okay to drink as a firefighter, but shameful to admit to AUD.
This prevents firefighters from seeking help for their AUD, or even admitting it, for fear of potentially ruining their career. Yet AUD poses a serious risk to individual firefighters and their teams, contributing to dangerous behaviors such as driving while intoxicated, heightened suicidal thoughts or actions, and an overall greater likelihood of traumatic incidents.
Currently, AUD screening relies on self-reporting through questionnaires. However, participants fear social and career-based repercussions or loss of their image if they admit to having an AUD. Objective screening methods are thus preferable, and those that do not stigmatize the participant are particularly beneficial. Behavioral and biological markers, such as structural MRI and neuropsychological assessments, are promising tools that form part of the present study.
This multimodal model was designed to use biological data to screen firefighters for high AUD risk.
Mapping risk through data
This South Korean study analyzed structural MRI neuroimaging data, combined with standardized neuropsychological tests, from a national cohort of 689 active-duty firefighters. This makes it the second-largest study of firefighters for AUD prediction modeling. The mean age was 43 years, with most participants being male.
The firefighters underwent T1-weighted structural MRI imaging. They also completed the Grooved Pegboard Test to assess their visual-motor coordination and the Trail Making Test to evaluate their executive function. They completed the Alcohol Use Disorder Identification Test (AUDIT), developed by the World Health Organization (WHO) to screen for alcohol dependence and addiction.
Participants were stratified into those with alcohol risk and non-alcohol risk, comprising 57% and 43%, respectively.
The study used a novel machine learning model that leveraged the synergistic benefits of acquiring different types of data. This included:
- ResNet-50 convolutional neural networks that extracted brain morphological patterns, layer by layer
- Vision Transformer modules to identify broad-based brain anatomy, relating different parts to each other
- Clinical variables used to derive a multilayer perceptron, identifying patterns in numerical data.
This multimodal system, known as “cooperative fusion,” integrates clinical and imaging data to interpret them in a clinically relevant manner. It combined brain patterns and clinical variations through their interactions, allowing for a more accurate prediction of AUD risk.
AI spots early warning signs
The multimodal system classified firefighters at risk for AUD with ~80% accuracy. Its discriminative power was also 80%. This represents a 17-percentage-point improvement in AUD prediction over either clinical-only or neuroimaging-only, at ~62% each.
The synergistic interaction of these different types of data accounts for the enhanced accuracy, showing that it is not simply an addition of more data. The model captured localized and global structural patterns from MRI data and integrated them effectively, eliminating the need for more complex and computationally demanding functional MRI (fMRI) scanning.
The neuropsychological tests also provided a satisfactory substitute for neurological function, further compensating for the lack of fMRI.
Using neuroimaging alone led to random activation patterns, indicating that early AUD produces too subtle and diffuse changes to provide enough discrimination to classify AUD risk accurately on its own.
Conversely, using only clinical features led to the use of sex and motor coordination as the most important predictive markers. The Grooved Pegboard test revealed sensitive changes in the nondominant hand, indicating early signs of neurological damage with AUD.
Model interpretability analyses such as Gradient-weighted Class Activation Mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP) further showed that unimodal imaging models produced diffuse, non-specific activation patterns. In contrast, the integrated multimodal model identified clear, biologically meaningful patterns, highlighting sex and motor coordination as key predictive features.
Again, the results emphasize the importance of sex-specific risk patterns for AUD in firefighters. Alcohol breakdown pathways, alcohol neurotoxicity, and addiction risk are known to vary by sex.
The researchers also evaluated model calibration and decision curve analyses, showing that the multimodal approach maintained both accuracy and clinical value across a wide range of threshold probabilities. It consistently outperformed single-modality models by reducing false positives and false negatives, whereas the accuracy of clinical-only and neuroimaging-only models declined at higher thresholds.
This indicates that this relatively simple machine learning model is relevant for clinical use. This is an important step forward, considering that the complexity of fMRI has been a drawback to the clinical applicability of multimodal approaches. It compares favorably with earlier research combining structural and functional MRI to classify participants by psychiatric diagnosis.
A safer path to prevention
The current study presents a novel multimodal machine learning framework that combines structural neuroimaging with neuropsychological tests of AUD-specific function. This provided better classification performance compared to either clinical-only or neuroimaging-only protocols. This reduces computational resources and image acquisition time.
The multimodal ResNet-50 + ViT + MLP model significantly outperformed both clinical-only and image-only models.
Larger studies are required to refine this approach, thereby increasing accuracy and preventing unnecessary career interventions. At present, it is best suited to current rather than prospective risk prediction, in view of the 20% misclassification rate.
Cost-effectiveness studies are also required. At present, it would require 150 screens to prevent one severe occupational incident at cost-effective levels. This does not account for cascading effects, such as the prevention of occupational risks and injuries, increased safety, lower liability exposure, and higher productivity and operational readiness.
However, this offers “a pragmatic pathway for implementing objective AUD screening in high-risk occupational populations with broader implications for psychiatric risk stratification in trauma-exposed professions.” Shorter screening protocols could be used during routine medical testing to maintain accuracy while improving feasibility and applicability, beyond firefighters to other high-risk professionals.
Download your PDF copy now!
Study links short-term blood pressure variability to Alzheimer's-related brain loss