This study utilized a cross-sectional design to develop and evaluate a multimodal deep learning framework for predicting alcohol use disorder (AUD) risk in an occupational cohort of active-duty firefighters in the Republic of Korea. Participants were recruited from multiple fire stations nationwide. Eligibility criteria included: age 25–65 years, active employment as a firefighter, and availability of both T1-weighted structural magnetic resonance imaging (MRI) and complete clinical assessment data. Exclusion criteria comprised a history of neurological disorders (e.g., epilepsy, stroke, traumatic brain injury), major psychiatric conditions other than AUD, current use of psychotropic medications, MRI-detected structural brain abnormalities, or contraindications to MRI scanning (e.g., metallic implants, claustrophobia).

Of 746 initially enrolled firefighters, 35 were excluded due to incomplete imaging data, 14 for missing clinical assessments, and 8 for MRI-detected structural anomalies, resulting in a final analytical sample of 689 participants (mean age 43.3 ± 8.8 years; 93% male). Figure 1 illustrates the participant recruitment and data preprocessing workflow. All participants provided written informed consent, and the study protocol was approved by the Institutional Review Board of Ewha Womans University. The research adhered to the ethical principles of the Declaration of Helsinki.

Cognitive and motor functions were evaluated using two standardized neuropsychological tests: the Grooved Pegboard Test and the Trail Making Test (TMT) (17, 18) (19, 20). The Grooved Pegboard Test assessed visual-motor coordination and fine motor control. Participants inserted 25 uniquely shaped pins into corresponding grooves as quickly as possible, with completion times (seconds) recorded for both dominant and non-dominant hands; longer times indicated poorer performance (17, 18). The TMT evaluated processing speed, cognitive flexibility, and executive function. Part A required participants to connect numbered circles sequentially, while Part B involved alternating between numbers and letters in ascending order. Completion times were recorded, with higher values reflecting lower cognitive efficiency (19, 20). Tests were administered by trained personnel under standardized conditions.

Alcohol use risk was assessed using the Alcohol Use Disorder Identification Test (AUDIT), a 10-item self-report questionnaire developed by the World Health Organization to evaluate alcohol consumption, dependence symptoms, and related harm (21). Scores range from 0 to 40, with a cutoff of ≥8 indicating hazardous drinking risk (22). The AUDIT was completed under supervised conditions to ensure data integrity.

To align comparisons with standard occupational screening practice, we stratified the cohort using the established AUDIT cut-off (≥8 vs<8). Table 1 summarizes the demographic and neuropsychological characteristics of the study cohort, stratified by AUDIT-based alcohol risk status (≥8: alcohol risk, n=392, 56.9%;<8: non-alcohol risk, n=297, 43.1%). The alcohol risk group had a mean age of 43.16 ± 8.53 years, compared to 42.58 ± 8.67 years for the non-alcohol risk group, with no significant difference (p=0.380, two-tailed independent samples t-test). We used two-tailed independent-samples t-tests for continuous variables because the groups are non-overlapping at the participant level, the t-test provides an efficient test of mean differences, and with our sample size it is reasonably robust to moderate deviations from normality; a two-sided test also guards against effects in either direction. A significant gender disparity was observed (p<0.001, two-tailed Pearson Chi-square test), with the alcohol risk group showing higher male predominance (380 males, 12 females) compared to the non-alcohol risk group (257 males, 40 females). The Pearson chi-square test was chosen for categorical comparisons (e.g., sex distribution) because it assesses association between group membership and categorical outcomes without requiring distributional assumptions beyond adequate expected cell counts.

Neuropsychological performance was comparable between groups. For the Grooved Pegboard Test, dominant hand completion times were 66.27 ± 8.45 seconds (alcohol risk) versus 67.27 ± 9.46 seconds (non-alcohol risk; p=0.153), and non-dominant hand times were 72.04 ± 9.49 seconds versus 72.48 ± 9.79 seconds (p=0.546). For the TMT, Part A completion times were 29.65 ± 7.78 seconds (alcohol risk) versus 28.71 ± 7.50 seconds (non-alcohol risk; p=0.110), and Part B times were 74.69 ± 27.28 seconds versus 73.42 ± 25.25 seconds (p=0.527). Interpreted under these method choices, the absence of significant between-group differences suggests that AUD risk, as defined by screening criteria, may precede measurable neuropsychological deficits in this occupational cohort.

Structural brain MRI scans were acquired using a 3.0 Tesla Philips MRI system (Philips Healthcare, Best, The Netherlands) equipped with a 32-channel head coil. High-resolution three-dimensional T1-weighted images were obtained with the following parameters: repetition time (TR) = 7.4 ms, echo time (TE) = 3.4 ms, flip angle = 8°, voxel size = 1 × 1 × 1 mm³, and 180 sagittal slices. All participants were instructed to maintain stillness and neutral head positioning throughout the scanning session to ensure image quality.

T1-weighted MRI data were preprocessed using the FMRIB Software Library (FSL, version 6.0 (23) to ensure standardized spatial normalization and artifact minimization. The preprocessing pipeline included both linear and nonlinear registration of each participant’s structural MRI to the Montreal Neurological Institute (MNI152) standard space, followed by resampling to a voxel resolution of 2 × 2 × 2 mm³. Following normalization, skull stripping was performed using the High-Definition Brain Extraction Tool (HD-BET), a deep learning–based algorithm designed to enhance the accuracy of brain tissue isolation from non-brain elements (24). This process improved visualization of key anatomical regions, including gray matter, white matter, and ventricular structures, while simultaneously reducing noise and enhancing segmentation fidelity. After skull stripping, the three-dimensional MRI volumes were segmented into 80 two-dimensional axial slices per participant, ensuring standardized anatomical coverage. The use of axial slices provides distinct spatial perspectives and clinically relevant information, facilitating detailed neuroanatomical interpretation and enabling precise detection of structural abnormalities or pathologies (25, 26).

To increase model generalizability, data augmentation techniques were applied. RandomAffine transformations introduced rotational variations (± 10°) and translation shifts (± 5%). ColorJitter transformations adjusted image brightness and contrast (± 20%), and RandomRotation transformations applied further rotational variation (± 15°) to simulate clinical variability. Pixel intensity normalization was performed using standardized mean and standard deviation values derived from large-scale neuroimaging datasets to standardize input distributions prior to model training (27–29).

Clinical assessment data underwent systematic preprocessing to ensure data integrity and model compatibility. Missing value analysis was conducted across all clinical variables, including demographic parameters (age, sex), neuropsychological test scores (Grooved Pegboard Test completion times for dominant and non-dominant hands, Trail Making Test Parts A and B), and alcohol use risk indicators (AUDIT scores). Participants with incomplete clinical assessments were excluded from the analytical cohort through listwise deletion, maintaining the methodological rigor requisite for multimodal integration. This conservative approach to missing data management, while potentially reducing statistical power, preserved the validity of cross-modal feature relationships critical to the multimodal learning framework. No imputation strategies were employed to avoid introducing artificial correlations between neuroimaging and clinical features. All continuous clinical variables were retained in their original scales to preserve interpretability, with normalization performed internally within the deep learning architecture through batch normalization layers. Categorical variables, specifically biological sex, were encoded using binary representation (0 = male, 1 = female) consistent with standard practices in medical machine learning applications.

To predict alcohol use disorder (AUD) risk in firefighters, we developed a multimodal deep learning framework that integrates structural magnetic resonance imaging (MRI) with clinical and neuropsychological data. The framework comprised three parallel processing branches: a convolutional neural network (CNN) based on ResNet-50 for local morphological feature extraction from MRI images (30), a Vision Transformer (ViT) module for global contextual representation of neuroanatomical structures (31), and a multilayer perceptron (MLP) for incorporating clinical and neuropsychological variables (32). Figure 2 provides a schematic overview of the deep learning architecture, illustrating the parallel processing of MRI images and clinical variables through the ResNet-50, Vision Transformer, and MLP modules, the subsequent feature concatenation, and the final classification layer.

For the MRI input stream, 80 axial two-dimensional slices per participant were fed into a pretrained ResNet-50 model to derive hierarchical local features. The resulting feature maps underwent average pooling and flattening operations to produce compact representations. In parallel, the same MRI slices were input into the ViT module via patch-based linear embedding. The ViT extracted long-range spatial dependencies and global structural context across brain regions (33). This dual-path design allowed for the concurrent extraction of both local and global representations from the neuroimaging data.

Simultaneously, clinical and neuropsychological features comprising age, sex, AUDIT score, Grooved Pegboard Test completion times (dominant and non-dominant hand), and Trail Making Test A and B durations were input into an MLP consisting of two fully connected layers with ReLU activations, yielding latent clinical representations. The outputs from the ResNet-50, ViT, and MLP branches were concatenated into a unified feature vector, which was passed through a fully connected layer with a sigmoid activation function to generate a binary prediction of AUD risk. Model training was conducted using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and a maximum of 100 training epochs. Early stopping was applied with a patience threshold of 10 epochs based on validation loss. To mitigate overfitting and improve generalizability, dropout regularization (dropout rate = 0.5) was applied to fully connected layers, and batch normalization was incorporated after each convolutional block.

Model evaluation was performed using stratified three-fold cross-validation with participant-level data partitioning to ensure independence between training and validation sets. Performance was assessed based on accuracy, area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity. Statistical differences in AUROC between model configurations were evaluated using DeLong’s test (34). All models were implemented in PyTorch (v1.10) and trained on an NVIDIA RTX A6000 GPU.

Model performance was evaluated using stratified threefold cross-validation with participant-level data partitioning to ensure independence between training and validation sets. Performance metrics included accuracy, area under the receiver operating characteristic curve (AUROC), precision, and recall. The AUROC was the primary metric due to its robustness to class imbalance. Confidence intervals (95% CI) for AUROC were estimated via bootstrapping (1,000 iterations).

Comparative analyses assessed the multimodal model against unimodal models (MRI-only, clinical-only). Between-model differences in AUROC were tested using DeLong’s method, which accounts for the correlation inherent to paired ROC curves evaluated on the same cases (34). Calibration was evaluated with reliability (calibration) curves to assess agreement between predicted probabilities and observed outcomes, and decision curve analysis was used to quantify net clinical benefit across threshold probabilities relevant to occupational screening. All preprocessing statistics, any calibration fits, and threshold selection were performed within training folds only and applied to the corresponding validation folds to avoid information leakage. Feature importance was analyzed using integrated gradients to enhance interpretability by identifying influential neuroanatomical and clinical inputs contributing to predictions (35). Statistical analyses were conducted using Python (version 3.9), Scikit-learn (version 1.0), and SciPy (version 1.7).

Feature importance analysis of clinical variables was conducted using SHAP methodology with an XGBoost classifier (36) trained on clinical features comprising age, sex, AUDIT scores, Grooved Pegboard Test completion times, and Trail Making Test durations. SHAP values were computed using TreeExplainer, which leverages the tree structure for efficient Shapley value calculation (37). Global feature importance was quantified through mean absolute SHAP values across the cohort, providing interpretable measures of each variable’s contribution to risk prediction. Statistical significance of feature contributions was evaluated using permutation-based null hypothesis testing with multiple comparison correction.

Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to elucidate the spatial localization of discriminative neuroanatomical features contributing to alcohol use disorder risk classification (38). This interpretability methodology generates visual explanations by computing the gradient of the predicted class score with respect to the final convolutional layer activations, thereby identifying brain regions that maximally influence the classification decision. The analysis targeted the terminal convolutional layers of each architecture, which preserve spatial resolution while encoding high-level semantic features. The importance weights α_k ^c for each feature map k with respect to target class c were computed through gradient backpropagation:

where y^c denotes the class score, A^k _ij represents the activation at spatial location (i,j) in feature map k, and Z normalizes by spatial dimensions. The final class-discriminative localization map was generated through weighted combination of forward activation maps:

The resulting coarse-grained heatmaps underwent bilinear interpolation to match the original image resolution (224×224 pixels) and were superimposed on the corresponding MRI slices with a transparency coefficient of 0.6 to facilitate anatomical interpretation. Visualizations were generated for a randomly selected subset of 50 participants per risk category to assess spatial consistency of learned features. Dice similarity coefficients quantified the spatial overlap of activation patterns across participants, while occlusion sensitivity analysis validated the causal importance of identified regions by measuring classification confidence degradation upon masking the upper quintile of activation intensities. Given these implementation details, we briefly justify our choice of localization method. We selected Grad-CAM after considering alternative feature-localization techniques because it is class-discriminative, CNN-architecture agnostic, and computationally efficient for 2D multi-slice MRI. Unlike vanilla saliency, which is high-variance and visually noisy, Grad-CAM yields stable, coarse-to-mid-scale heatmaps aligned with the target class. Integrated Gradients requires a baseline and path integral whose choice is non-trivial for T1 intensity scales and can introduce baseline-dependent artifacts, whereas Grad-CAM avoids a baseline choice while remaining faithful to score–gradient information. Occlusion/perturbation and LIME/SHAP image explanations impose heavy sampling costs and design choices (e.g., patch size, superpixels) that scale poorly to ~80 slices per subject (37, 39). Transformer attention maps are not inherently class-specific and may not reflect decision-critical evidence, whereas Grad-CAM is explicitly class-discriminative. In medical imaging, Grad-CAM’s regional localization aligns with radiological reading practices, enabling transparent overlays on axial slices and cohort-level aggregation without re-training. To address known limitations (resolution tied to the last conv layer), we performed sanity checks (parameter randomization and slice-wise ablation) and report both representative and aggregated maps (40).

Table 3 provides a comprehensive summary of recent multimodal deep learning approaches for psychiatric disorder prediction, contextualizing our findings within the broader landscape of neuroimaging-based classification studies. Contemporary neuroimaging investigations have consistently demonstrated the superiority of multimodal approaches combining structural MRI, functional task-based MRI, and resting-state functional connectivity in psychiatric classification tasks (46). However, the multimodal framework presented herein, achieving 79.88% accuracy (AUROC: 0.795) through structural MRI and clinical assessment integration alone, demonstrates competitive performance relative to architectures incorporating functional neuroimaging. A recent triple-modality integration study (42) (sMRI + fMRI + SNP) yielded 79.01% accuracy in schizophrenia classification (n=492), with individual modalities contributing differentially (sMRI: 66.33%, fMRI: 75.29%, SNP: 57.06%). The marginal improvement from sMRI baseline to multimodal integration (13.68 percentage points) must be contextualized against substantially increased acquisition complexity and computational burden. A comprehensive investigation utilizing an extensive neuroimaging battery comprising 119 alcohol-dependent patients and 97 controls revealed that while multimodal integration yielded optimal classification performance, the investigators concluded that “in terms of direct clinical applicability, currently the most realistic neuroimaging-based classifier for AD may be unimodal based on structural MRI and grey-matter density specifically” (46), citing the temporal demands and analytical complexity of functional MRI protocols. This empirical observation corroborates our methodological decision to prioritize T1-weighted structural MRI as the primary neuroimaging modality.

Previous investigations employing resting-state functional connectivity have reported classification accuracies ranging from 61.53% to 76.67% for discriminating alcohol-dependent individuals from controls (43, 44). Direct comparison with AUD-focused investigations reveals our framework’s competitive performance despite methodological parsimony. A recent study (44)reported resting-state fMRI yielding AUROC 0.79 (n=102), comparable to our 0.795 despite utilizing computationally intensive connectivity analyses. This equivalence challenges assumptions regarding the superior discriminative capacity of functional imaging for AUD detection. Similarly, another investigation (46) demonstrated that dual neuroimaging modality integration (sMRI + fMRI) achieved 79.3% accuracy in AUD classification (n=216), representing merely 2.7 percentage points improvement over single modality (76.6%) which represents a limited enhancement that raises critical questions regarding the cost-effectiveness of functional imaging protocols in occupational screening contexts. Random Forest classification leveraging functional connectivity within the Default Mode Network combined with neuropsychological measures achieved 76.67% accuracy (45), necessitating extensive preprocessing pipelines and network-level analytical frameworks. Recent work (45) reported fMRI-based classification achieving 76.67% accuracy (AUROC: 0.93) in a male-exclusive cohort (n=60). Our superior accuracy (79.88%) in a substantially larger sample (n=689) with mixed-gender composition suggests that structural alterations combined with behavioral assessments may provide greater discriminative capacity than functional connectivity alone in occupational populations. Resting-state connectivity features have demonstrated capacity to explain 33% of variance in Alcohol Use Disorders Identification Test (AUDIT) scores (47), though such models required acquisition of multiple functional MRI sequences including monetary incentive delay and face-matching paradigms alongside resting-state protocols.

Multimodal data integration approaches in psychiatric research demonstrate methodological advantages comparable to our neuroimaging-clinical framework. A recent VR-based study (48) developed machine learning models utilizing acoustic and physiological features VR exposure sessions for social anxiety disorder, achieving an AUROC of 0.852 with CatBoost for Social Phobia Scale prediction using multimodal features (n=132 samples from 25 participants). Notably, their analysis revealed that acoustic features (AUROC: 0.788) substantially outperformed physiological features alone (AUROC: 0.626) for anxiety symptom prediction, with multimodal integration yielding superior classification performance across multiple anxiety domains. While their VR-based approach differs methodologically from our structural neuroimaging framework, the 7.26 percentage point improvement from physiological to multimodal features (compared to our 17.35 percentage point improvement from clinical to multimodal) highlights the consistent benefit of cross-modal integration in psychiatric risk stratification. Their findings that acoustic biomarkers captured more discriminative information than physiological responses during anxiety-inducing scenarios parallels our observation that targeted neuropsychological assessments provide critical functional anchoring for structural alterations.

The classification performance of our T1-weighted structural MRI multimodal approach (79.88% accuracy) demonstrates favorable comparison with functional connectivity-based methodologies while offering considerable practical advantages regarding acquisition efficiency and computational parsimony. Our framework’s 17.35 percentage point improvement from clinical-only baseline (62.53%) substantially exceeds the incremental gains observed when adding neuroimaging to clinical data reported previously (46), suggesting that targeted neuropsychological assessments may capture variance typically attributed to functional connectivity measures. A systematic review examining machine learning applications in AUD reported neuroimaging-based algorithms achieving sensitivity ranging from 90-99.99% and specificity from 82-99.97% (14); however, these exceptional performance metrics were predominantly observed in investigations combining multiple imaging modalities. One study (43)reported 3D sMRI + fMRI combination achieving accuracy ranging from 67.4% to 90.5% (n=92). The substantial variability suggests potential overfitting in small samples, emphasizing the importance of our larger cohort (n=689) for robust generalization estimates. Beyond structural neuroimaging approaches, recent advances in machine learning applications for mental health monitoring in first responders provide complementary perspectives on psychological distress prediction. A proof-of-concept investigation (49) developed predictive models for posttraumatic stress injuries (PTSI) utilizing intensive longitudinal data from 274 Montreal firefighters monitored biweekly across 12 weeks. The study implemented four distinct machine learning algorithms (logistic regression, support vector classifier, extreme gradient boosting) trained on temporal sequences of standardized psychological assessments (PHQ-9, GAD-7, PCL-5) and psychosocial variables (occupational stress, social support, coping strategies). The optimal model configuration, employing extreme gradient boosting with three lagged measurement timepoints and comprehensive feature sets, achieved 94% classification accuracy (AUC = 0.93, sensitivity = 0.61, specificity = 0.97). Several methodological contrasts with the present investigation merit consideration. The documented PTSI prevalence, fluctuating between 6.9% and 10.6% across assessment intervals with cumulative incidence of 19.7%, represents substantially lower psychopathology rates than our observed AUD risk prevalence of 56.9%, potentially attributable to differential diagnostic thresholds between acute stress-related symptomatology and chronic alcohol use vulnerability. Feature importance analyses identified lagged PHQ-9 scores collected 2 and 6 weeks prior to target assessment as dominant predictors (19% and 10% relative importance respectively), with GAD-7 and PCL-5 scores contributing secondarily, while demographic variables (age >46 years, work experience >21 years) demonstrated minimal predictive value. This hierarchical pattern corresponds with our SHAP-derived feature attributions wherein neuropsychological performance metrics superseded demographic characteristics. The temporal dependency of predictive accuracy, wherein model performance systematically improved from single-timepoint (accuracy range: 0.81-0.91) to three-timepoint configurations (accuracy range: 0.82-0.94), underscores the critical importance of longitudinal symptom trajectories in psychiatric risk modeling. These convergent findings across distinct methodological paradigms substantiate the superiority of multimodal, temporally-informed approaches over cross-sectional univariate assessments. Integration of periodic structural neuroimaging for baseline vulnerability characterization with continuous smartphone-based symptom monitoring could potentially optimize early intervention strategies through synthesis of stable neurobiological markers and dynamic clinical trajectories. Our approach achieves minimal sacrifice in predictive accuracy while greatly reducing acquisition time, computational burden, and technical expertise requirements (49).

Previous investigations utilizing isolated structural MRI modalities have provided valuable performance benchmarks, with grey matter density analysis achieving 65% classification accuracy in comprehensive multimodal comparisons (46). The present study builds upon these findings by demonstrating that augmenting structural neuroimaging with targeted neuropsychological assessments yields enhanced discriminative capacity (79.88% accuracy), consistent with theoretical frameworks positing synergistic information capture across neurobiological and behavioral domains. The 17.35 percentage point improvement from clinical baseline reflects fundamental complementarity rather than simple feature concatenation: structural neuroimaging captures cumulative morphological alterations reflecting chronic alcohol exposure, providing stable biomarkers of neurotoxic burden, while neuropsychological performance offers dynamic functional readouts of neural system integrity sensitive to subclinical impairments. This performance differential underscores the critical importance of incorporating standardized neuropsychological assessments to compensate for the absence of functional connectivity information. Recent investigations have emphasized that machine learning algorithms provide valuable tools for quantifying large-scale network differences in AUD (44); however, our results suggest that morphological features combined with targeted clinical assessments achieve comparable discriminative capacity.

The firefighter population presents unique challenges for AUD prediction modeling. While general population studies have examined heterogeneous samples characterized by varied substance use histories and psychiatric comorbidities (43), our cohort’s occupational homogeneity and elevated baseline risk necessitated tailored analytical approaches. With n=689, our investigation represents the second-largest cohort among reviewed studies [following a recent MDD study (41): n=2319], providing robust statistical power while maintaining occupational homogeneity. While T1-weighted structural sequences and fMRI share similar acquisition times (5–10 minutes each), the critical distinction lies in post-processing complexity. Functional MRI necessitates sophisticated preprocessing pipelines encompassing motion correction, temporal filtering, spatial smoothing, and connectivity analysis, extending analysis time from hours to days. Additionally, fMRI’s heightened motion sensitivity increases data attrition rates, compromising practicality for large-scale screening initiatives. Previous occupational cohort investigations remain limited, constraining direct performance comparisons. Nevertheless, the effective classification achieved without functional MRI suggests that structural alterations and behavioral manifestations may exhibit enhanced discriminability in high-risk occupational groups, potentially attributable to chronic stress exposure and cultural factors influencing alcohol consumption patterns. The elimination of fMRI-specific infrastructure requirements (stimulus presentation systems, synchronization hardware, specialized preprocessing software) substantially reduces implementation barriers in clinical settings, supporting the translational feasibility of our approach for occupational health surveillance.

The efficacy of our multimodal approach necessitates examination through complementary interpretability methodologies to elucidate differential contributions of neuroimaging and clinical features. Recent advances in multimodal explainable AI have demonstrated the critical importance of understanding feature interactions across modalities. A recent study achieved 94.81% accuracy using an Ensemble Optimization-enabled Explainable CNN (EO-ECNN) with multimodal data integration, highlighting the significance of interpretability in clinical applications (50). Recent comparative studies have systematically evaluated various explainability approaches for multimodal medical imaging. A large-scale experiment across four medical imaging datasets found that while attention maps from Vision Transformers generally surpass Grad-CAM in explainability, transformer-specific interpretability methods demonstrate superior performance (51). This finding underscores the importance of selecting architecture-appropriate interpretability techniques rather than applying traditional CNN-based methods to transformer architecture. Gradient-weighted Class Activation Mapping (Grad-CAM) analysis applied to unimodal neuroimaging models revealed critical insights regarding the limitations of structural MRI-only approaches. As illustrated in Supplementary Figure S1 , Grad-CAM visualizations from both ResNet-50 (Panel A) and EfficientNet-B0 (Panel B) architectures demonstrated stochastic activation patterns across randomly selected axial brain slices. The activation maps exhibited no systematic concentration within anatomically relevant regions associated with alcohol-related neurodegeneration, instead displaying diffuse, heterogeneous patterns with focal hotspots appearing randomly across cortical and subcortical territories. The stochastic activation patterns observed through Grad-CAM analysis provide empirical evidence for the fundamental limitations of structural MRI-only approaches in detecting subtle, distributed alterations associated with AUD risk. This finding aligns with previous neuroimaging studies showing that morphological changes in early-stage AUD are often diffuse and heterogeneous, requiring behavioral anchoring for meaningful interpretation (46, 52).

This absence of neuroanatomically coherent feature extraction in image-only models provides mechanistic validation for observed performance limitations (ResNet-50: 57.73% AUROC; EfficientNet-B0: 56.54% AUROC). A comprehensive survey of explainable multimodal learning methods confirmed that such random activation patterns indicate insufficient discriminative capacity when structural alterations are subtle and distributed (53). Recent advances in transformer architecture have introduced attention visualization as a complementary interpretability approach. Studies on multimodal foundation models for anomaly detection have demonstrated that combining SHAP, Grad-CAM, and attention visualization provides more comprehensive insights than any single approach, particularly when dealing with heterogeneous medical data sources (54). These findings suggest that different XAI techniques capture complementary aspects of model behavior: spatial localization through Grad-CAM, feature importance through SHAP, and hierarchical relationships through attention mechanisms. Grad-CAM heatmaps revealed that convolutional neural networks, when constrained to structural MRI data alone, failed to converge on consistent morphological markers despite well-established volumetric alterations in alcohol-dependent populations. Peak activation intensities varied markedly between slices without correspondence to regions of established vulnerability including prefrontal cortex, hippocampus, or cerebellar structures. This stochastic behavior suggests that structural alterations alone, while present, may be insufficiently discriminative for effective classification without complementary functional or behavioral indicators.

Conversely, SHapley Additive exPlanations (SHAP) analysis of clinical variables within the optimal multimodal framework revealed hierarchical feature importance with clear mechanistic interpretability ( Supplementary Figure S2 ). Sex emerged as the predominant contributor with SHAP values ranging from -0.05 to +0.45, exhibiting pronounced rightward skew indicative of male sex as a risk amplifier. This finding aligns with established sex differences in alcohol metabolism, neurotoxic vulnerability, and addiction trajectories. The prominence of biological sex as a predictor (SHAP values: -0.05 to +0.45) aligns with established epidemiological evidence showing higher AUD prevalence in male firefighters and known sex differences in alcohol metabolism and neurotoxic vulnerability (55). Non-dominant hand Grooved Pegboard performance demonstrated bidirectional effects (SHAP values: -0.30 to +0.25), suggesting that motor coordination impairments serve as sensitive indicators of subclinical neurological compromise.

The substantial performance differential between the multimodal framework (79.88% accuracy) and both clinical-only (62.53%) and imaging-only (61.53%) approaches cannot be attributed to simple additive effects of feature concatenation. Rather, this performance enhancement reflects fundamental principles of multimodal machine learning wherein complementary information sources capture distinct aspects of underlying pathophysiology (56, 57). Recent theoretical frameworks in multimodal neuroimaging emphasize that different data modalities provide non-redundant views of complex biological phenomena, with optimal integration strategies exploiting this complementarity to achieve superior discriminative power (58, 59).

The integration of neuroimaging and clinical features within our multimodal architecture leverages what has been termed “cooperative fusion” in the multimodal learning literature, wherein modalities interact synergistically to reveal patterns invisible to either modality in isolation (60). This approach aligns with recent comprehensive reviews demonstrating that transformer and CNN architectures require tailored interpretability methods to effectively capture their distinct feature extraction patterns (61, 62). Structural MRI captures static morphological alterations reflecting cumulative neurotoxic effects, while neuropsychological assessments provide dynamic functional readouts of neural system integrity. The Vision Transformer component, designed to capture global spatial dependencies, may identify distributed patterns of subtle atrophy that achieve diagnostic relevance only when contextualized by concurrent functional deficits captured through clinical assessments. This synergistic relationship aligns with recent findings demonstrating that multimodal approaches consistently outperform unimodal methods in neuropsychiatric classification tasks by capitalizing on the complementary nature of structural and functional information (63, 64).

Critically, observed performance gains cannot be explained by overfitting to clinical features or trivial demographic correlations. SHAP analysis reveals that while sex contributes significantly, motor coordination measures and other neuropsychological indicators provide substantial independent predictive value. Moreover, the failure of clinical-only models to exceed 62.53% accuracy demonstrates that behavioral assessments alone lack sufficient discriminative capacity. Similarly, poor performance of imaging-only models indicates that structural alterations, though present, require behavioral anchoring for meaningful interpretation in this at-risk but pre-clinical population.

Mechanistic insights derived from interpretability analyses have profound implications for understanding AUD vulnerability in occupational cohorts. The failure of image-only models to identify consistent neuroanatomical markers suggests that structural alterations in early-stage or at-risk individuals may be subtle, distributed, and heterogeneous, requiring behavioral anchors for meaningful interpretation. The prominence of sex and motor coordination measures in feature importance rankings indicates that integrative models capturing both biological predisposition and functional manifestation provide superior discriminative capacity. These findings support a multifactorial conceptualization of AUD risk wherein neurobiological alterations interact with demographic vulnerabilities and manifest through measurable performance deficits before clinical thresholds are reached.

This investigation presents several methodological constraints warranting comprehensive examination. First, the cross-sectional design fundamentally precludes causal inference regarding temporal evolution of structural brain alterations and their relationship to AUD risk. Longitudinal investigations tracking firefighters from recruitment through career progression would be essential to establish whether observed neuroanatomical variations represent predisposing vulnerabilities, consequences of occupational stress exposure, early markers of problematic alcohol use, or complex interactions among these factors. The absence of temporal data particularly limits our ability to determine whether structural alterations precede behavioral manifestations or emerge concurrently with escalating alcohol consumption.

Additionally, exclusive reliance on T1-weighted structural MRI, while strategically chosen for clinical feasibility, imposes inherent constraints on the comprehensiveness of neurobiological characterization. Resting-state functional MRI investigations have identified specific functional connectivity alterations in reward, salience, and executive control networks that differentiate individuals with AUD from controls (43, 44, 47). Our structural-only approach cannot capture these dynamic network-level disruptions, potentially missing critical neurophysiological markers of addiction vulnerability. Future iterations incorporating abbreviated resting-state protocols or task-based functional MRI targeting reward processing could enhance predictive accuracy while maintaining reasonable clinical practicality. Additionally, advanced structural imaging techniques such as diffusion tensor imaging could provide microstructural integrity measures complementing volumetric assessments.

Furthermore, the computational decision to segment three-dimensional brain volumes into two-dimensional axial slices, while reducing computational complexity and memory requirements, sacrifices spatial continuity information. Three-dimensional convolutional architectures or graph neural networks operating on whole-brain volumes could potentially capture long-range anatomical relationships and improve feature extraction, though at substantially increased computational cost. The trade-off between model sophistication and practical deployability remains a critical consideration for clinical translation.

Moreover, generalizability of our findings faces multiple constraints. The sample comprised exclusively Korean firefighters, introducing both cultural and occupational specificity that may limit applicability to other populations. Cultural variations in alcohol consumption patterns, stigma associated with help-seeking, and occupational stress exposure could influence both structural brain alterations and model performance. The marked sex imbalance (93% male) reflects firefighting workforce demographics but severely limits conclusions about female firefighters, particularly given sex differences in alcohol metabolism, vulnerability to neurotoxic effects, and addiction trajectories. Validation in diverse cultural contexts, occupational groups, and sex-balanced samples remains essential before broader implementation.

Concomitantly, several technical limitations merit consideration. The AUDIT threshold of 8 for defining at-risk status, while internationally validated, may not optimally discriminate problematic drinking patterns in high-functioning occupational cohorts where normative drinking levels differ from general populations. The relatively modest sample size (n=689), while substantial for neuroimaging studies, may limit detection of subtle subgroup differences or complex interaction effects. The absence of genetic data precludes investigation of gene-environment interactions known to influence AUD vulnerability. Missing longitudinal follow-up data prevents validation of the model’s actual predictive utility for incident AUD diagnosis or occupational impairment.

Beyond these methodological considerations, clinical implementation faces practical challenges beyond model performance. The requirement for high-resolution structural MRI limits deployment to settings with advanced imaging facilities, potentially excluding rural or resource-limited fire departments. The need for standardized neuropsychological testing by trained personnel adds operational complexity. Privacy concerns regarding neuroimaging-based occupational screening require careful ethical consideration and policy development. The potential for algorithmic bias, particularly given the homogeneous training sample, necessitates ongoing monitoring and recalibration in diverse deployment contexts.

Finally, performance ceiling effects warrant critical consideration. While the achieved 79.88% accuracy represents competitive performance relative to existing multimodal approaches, it nonetheless implies a 20.12% misclassification rate with asymmetric consequences: false positives potentially triggering unwarranted career interventions versus false negatives resulting in missed opportunities for early therapeutic engagement. Furthermore, the absence of longitudinal follow-up data fundamentally constrains clinical interpretation; the model’s capacity to predict incident AUD diagnoses, trajectory of symptom progression, or subsequent occupational impairment remains empirically undefined. This temporal limitation restricts current applicability to cross-sectional risk stratification rather than prospective prediction, highlighting the imperative for longitudinal validation studies to establish true predictive validity and optimal screening intervals.

The demonstrated feasibility of achieving robust AUD risk classification using structural MRI represents a significant advancement for occupational health surveillance in high-risk professions. Traditional screening paradigms relying on self-report instruments face systematic limitations in emergency response populations where cultural valorization of stoicism, occupational stigma, and career preservation concerns suppress accurate disclosure of alcohol-related problems (2, 12). The integration of objective neurobiological markers derived from structural neuroimaging circumvents these reporting biases while maintaining classification performance comparable to more complex multimodal approaches.

Implementation within existing occupational health frameworks requires consideration of both technical infrastructure and organizational factors. Fire departments conducting periodic comprehensive medical evaluations could incorporate T1-weighted structural MRI protocols, though the total examination time of 30–60 minutes represents a substantial logistical consideration. While the T1-weighted sequence itself requires only 5–10 minutes of actual acquisition time, the complete imaging protocol including patient preparation, positioning, and safety procedures necessitates dedicated scheduling within occupational health assessments. The demonstrated predictive value of combining neuroimaging with standardized neuropsychological assessments suggests that brief cognitive testing batteries could enhance screening accuracy without requiring specialized neuropsychological expertise. Automated analysis pipelines utilizing the validated deep learning architecture could provide rapid risk stratification post-acquisition, enabling occupational health physicians to prioritize intervention resources toward highest-risk individuals.

To address feasibility concerns, a tiered implementation strategy could optimize resource utilization while maintaining screening effectiveness. Initial deployment could target high-risk subpopulations identified through traditional screening tools (AUDIT scores ≥ 15) or those with documented occupational incidents, thereby concentrating MRI resources on individuals with greatest clinical need. As infrastructure develops and costs decrease, screening criteria could progressively expand to encompass broader firefighter populations. This phased approach aligns with successful precedents in occupational health screening, where targeted protocols for high-risk workers preceded universal implementation (65).

Critical ethical considerations must guide translation from research findings to occupational screening practices. Clear delineation between probabilistic risk assessment and clinical diagnosis remains essential to prevent discriminatory practices while maximizing preventive potential. Screening protocols should emphasize early intervention and support rather than punitive measures, with explicit protections ensuring that neuroimaging findings cannot adversely impact employment status without corroborating clinical evidence. Longitudinal monitoring frameworks tracking predictive validity of initial risk assessments against subsequent clinical outcomes would enable continuous algorithm refinement while building evidence for screening effectiveness.

Economic implications of neuroimaging-based screening warrant careful analysis within resource allocation frameworks. Recent economic modeling of MRI-based screening programs provides relevant benchmarks. The UK Biobank’s population neuroimaging initiative achieved per-scan costs of £264 ($330 USD) through high-volume standardization (66). Considering firefighters’ elevated AUD risk (56.9% in our cohort vs. 6.2% general population) (67) and associated costs of untreated AUD (estimated $249 billion annually in the US) (68), targeted neuroimaging screening may prove cost-effective despite initial infrastructure investments. A threshold analysis suggests that preventing one severe occupational incident per 150 screenings would offset implementation costs. Furthermore, downstream savings from prevented occupational injuries, reduced absenteeism, decreased liability exposure, and maintained operational readiness strengthen the economic justification. Insurance frameworks may require modification to recognize preventive neuroimaging in high-risk occupational cohorts as medically necessary, particularly given these demonstrated cost-benefit ratios. Future development priorities should address current limitations while enhancing clinical utility. Multicenter validation studies incorporating diverse geographical regions, departmental cultures, and demographic compositions would establish generalizability boundaries and identify population-specific calibration requirements. Integration with emerging digital biomarkers from wearable devices, sleep monitoring, and stress physiology could create comprehensive risk profiles extending beyond cross-sectional neuroimaging snapshots. Development of abbreviated screening protocols optimized for rapid deployment during routine medical evaluations could enhance feasibility while maintaining predictive accuracy.

The broader implications extend beyond firefighting to encompass other high-stress occupations with elevated substance use risk, including law enforcement, emergency medical services, and military personnel. Establishing standardized neuroimaging protocols and classification algorithms across these populations would enable comparative effectiveness research while building robust normative databases. International collaboration through occupational health networks could accelerate validation efforts while ensuring equitable access to advanced screening technologies across resource-varied settings.