A machine learning-based screening tool for classifying post-traumatic stress disorder in firefighters: Firefighter Latent Risk Evaluation (FLARE) Authors: Seok Joo Chae, Johanna Inhyang Kim, Hyeontae Jo, Yun Min Song, Saebom Jeon, Sun-Young Moon, Ji-Hye Lee, Sohee Oh, Seockhoon Chung, Jae Kyoung Kim, Jeong-Hyun Kim Published: 2026 ABSTRACT AbstractBackgroundThe prevalence of post-traumatic stress disorder (PTSD) among South Korean firefighters is likely to be under-reported because of stigma and defensive attitudes. We aimed to develop a machine learning-based classification model, named Firefighter PTSD Latent Risk Evaluation (FLARE), to identify firefighters at risk of probable PTSD based on demographic and clinically relevant psychosocial features.MethodsWe analyzed data from 52,428 firefighters, including 2,305 with probable PTSD and 50,123 without PTSD, as assessed using the PTSD Checklist for DSM-5. Various clinical and occupational factors were examined, including exposure to traumatic events, depression, alcohol use disorder, insomnia, occupational stress, resilience, and awareness of and attitude towards mental illness. An extreme gradient boosting (XGBoost) algorithm was applied using the selected features to develop a FLARE model for PTSD classification. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC).ResultsThe FLARE model, comprising seven key features (decreased concentration, appetite changes, feelings of worthlessness or excessive/inappropriate guilt, sleep disturbances, perceived work pressure, poor adaptability to change, and low perceived appropriateness of seeking professional help for close others) achieved an AUROC of 0.90 in classifying probable PTSD, with a sensitivity of 0.84 and a specificity of 0.82 in the test dataset. Temporal validation using the 2024 dataset yielded an AUROC of 0.91, with a sensitivity of 0.83 and a specificity of 0.84, confirming the model’s generalizability.ConclusionsThe FLARE model effectively identified firefighters at risk for PTSD using a concise set of clinical and psychosocial indicators. This tool has the potential to enhance PTSD screening and early interventions, particularly in populations with prevalent mental health stigma.Clinical trial numberNot applicable.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12888-026-08151-5.Keywords: PTSD, Firefighters, Classification, Machine learning, Stigma, Occupational stress Background Firefighters are regularly exposed to traumatic events, placing them at high risk of developing post-traumatic stress disorder (PTSD) throughout their careers [1, 2]. Previous studies have reported that 31.8% of U.S. firefighters exhibit probable PTSD [3], whereas a global meta-analysis estimated a pooled prevalence of 10.0% among firefighters and rescue workers [4]. However, national studies on South Korean firefighters have consistently reported significantly lower PTSD rates of 2.6% in 2019 [5] and 2.9% in 2020, based on self-reported questionnaires [6]. This discrepancy may be partially attributed to pervasive concerns about stigma with PTSD often perceived as “weakness” or “vulnerability” [7]. Moreover, social desirability bias, which is particularly strong in East Asian cultures, may contribute to underreporting [8]. South Korean firefighters are legally required to participate in annual mental health screenings (Article 16 of the Special Health Checkups for Fire Officials). However, repeated exposure to PTSD screening questionnaires may have led to increased familiarity with the PTSD diagnostic criteria, potentially resulting in defensive responses and reduced reliability of self-reported PTSD assessments. This could further contribute to the underreporting of PTSD in this population. Rather than relying solely on direct PTSD symptom assessments, this study proposed a novel approach: identifying PTSD risk through clinically relevant psychosocial symptoms and demographic features commonly reported by firefighters with PTSD. This method may provide a more reliable means of detecting underreported probable PTSD and highlight the need for further clinical evaluation. Previous studies identified various clinical, demographic, and occupational factors associated with PTSD among firefighters. A recent review of PTSD determinants among military personnel and firefighters highlighted multiple contributing factors including demographic characteristics, organizational stressors, psychiatric comorbidities, social support, and personality traits [9]. Additionally, psychiatric conditions such as depression, insomnia, and alcohol use disorder (AUD) frequently co-occur with PTSD in firefighters [10], suggesting a bidirectional relationship; preexisting psychiatric disorders may increase the risk of PTSD, whereas PTSD can exacerbate these conditions. Firefighters experience high occupational stress owing to physical demands, heavy workloads, frequent emergency calls, and irregular work hours [11]. Occupational stress has been linked to negative mental health outcomes, including depression and PTSD [1, 2]. Moreover, the stigma surrounding mental health issues, often perceived as a sign of incompetence or weakness, can exacerbate PTSD prevalence by increasing anxiety and hindering help-seeking behaviors [12–14]. Firefighters in particular experience significant pressure to maintain a strong persona, and stigma has been shown to correlate with increased PTSD prevalence by discouraging social support and professional mental health care [15]. However, research on the collective influence of these factors on the risk of developing PTSD among firefighters remains limited. Recent advancements in machine learning have significantly improved the analysis of clinical questionnaires by identifying key predictive features and enhancing the diagnostic accuracy. Machine learning algorithms have been successfully applied to assess sleep disorders, sleep-related cognition, and insomnia indices, demonstrating their ability to extract critical features and simplify screening tools, while enhancing accuracy [16–20]. By leveraging these algorithms to isolate the most significant items, researchers can refine long-term clinical questionnaires into concise and effective classification models, thereby increasing their practicality for real-world applications. We aimed to address two key gaps in PTSD research among firefighters: the limited understanding of how multiple known risk factors interact to contribute to PTSD development and the potential limitations of self-reported PTSD questionnaires, which may be influenced by stigma, social desirability bias, or defensive response. Although clinician-administered diagnostic interviews remain the gold standard for the diagnosis of PTSD, self-reported measures are indispensable for large-scale mental health screening. This study leveraged machine learning to develop a classification model for probable PTSD, using questionnaire items that assessed the symptoms of various clinical, psychosocial and demographic factors associated with PTSD. Unlike previous studies that focused primarily on psychiatric diagnoses, this study identified key clinical symptoms that were easily accessible through self-report measures, ultimately creating a practical and efficient screening tool for PTSD in firefighters. Methods Data collection South Korean firefighters undergo mandatory annual mental health screening, as required by law (Article 16 of the Framework Act on the Health, Safety, and Welfare of Fire Officials). The National Fire Agency conducted a self-report online survey between April 7 and May 6, 2023, targeting firefighters from 215 fire stations across the country. Of 64,232 firefighters invited to participate, 52,802 (82.2%) completed the survey. An anonymized dataset was provided to the researchers for analysis. After excluding responses with missing or invalid data, 52,428 firefighters were included in the final analysis. For the temporal validation of the developed model, we additionally utilized data from the 2024 annual mental health screening. This survey was conducted from January 22 to March 3, 2024, targeting firefighters across the country, providing an anonymized dataset of 57,528 firefighters. Although a substantial overlap in participants between the 2023 and 2024 datasets is expected given the nature of the mandatory annual screening, the data were completely anonymized. Therefore, the 2024 dataset was treated as a cross-sectional cohort to evaluate the temporal stability and generalization of our model. Measures The survey collected data on the following variables: demographic and occupational characteristics (age, sex, years of work experience, and shiftwork schedules categorized into seven one-hot encoded types), exposure to traumatic experiences, PTSD symptoms, depression, alcohol use disorders, insomnia, occupational stress, resilience, and attitudes toward mental illness. PTSD checklist for DSM-5 (PCL-5) PTSD symptoms were assessed using the 20-item Korean version of the PTSD Checklist for DSM-5 (PCL-5), a widely used screening tool for PTSD [21]. Participants identified the most impactful traumatic event from a predefined list and rated each item on a 5-point Likert scale (0 = not at all, 4 = extremely). The PCL-5 assesses symptoms across four PTSD subdomains. Intrusion: At least one item rated 2 or higher from items 1–5. Intrusion: At least one item rated 2 or higher from items 1–5. Avoidance: At least one item rated 2 or higher from items 6–7. Avoidance: At least one item rated 2 or higher from items 6–7. Negative alterations in cognition and mood: At least two items were rated 2 or higher on items 8–14. Negative alterations in cognition and mood: At least two items were rated 2 or higher on items 8–14. Increased arousal: At least two items rated 2 or higher among items 15–20. Increased arousal: At least two items rated 2 or higher among items 15–20. Participants meeting all four symptom criteria were classified as having probable PTSD. Exposure to traumatic experiences The frequency of traumatic exposure over the past year was measured using an adapted version of the questionnaire by Beaton et al. [22]. Of the original 33 stressors, 24 stressors considered to be related with high levels of stress were selected, excluding two related to gun violence because of its low incidence in South Korea. Additionally, the survey included two modified items regarding witnessing or not witnessing the suicide of a co-worker and four new items on handling the remains of suicide victims, disposing of severely decayed corpses, involvement in high-profile public safety incidents, and experiencing workplace audits/inspections. A complete list of traumatic events is provided in Table S1. Patient health questionnaire-9 (PHQ-9) Depressive symptoms were assessed using the Korean version of the Patient Health Questionnaire-9 (PHQ-9) [23, 24]. The PHQ-9 consists of nine items reflecting the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV) criteria for major depressive disorder, rated on a 4-point Likert scale (0 = not at all, 3 = nearly every day), reflecting symptoms over the past 2 weeks. The total score ranges from 0 to 27, with higher scores indicating a greater severity of depression. 0–4: no depression. 0–4: no depression. 5–9: mild depression. 5–9: mild depression. 10–19: moderate to moderately severe depression. 10–19: moderate to moderately severe depression. 20–27: severe depression. 20–27: severe depression. In the present study, the Cronbach’s alpha for the PHQ-9 was 0.91, indicating excellent internal consistency. Alcohol use disorders identification test (AUDIT) The severity of alcohol use problems was assessed using the alcohol use disorders identification test (AUDIT), developed by the World Health Organization (WHO) [25, 26]. The AUDIT comprises 10 items, 3 of which assess drinking frequency, alcohol quantity, and binge drinking, whereas the remaining seven items evaluate alcohol-related consequences. Higher scores indicate greater alcohol-related risks. 1 to 7: low risk. 1 to 7: low risk. 8 to 14: hazardous or harmful drinking. 8 to 14: hazardous or harmful drinking. ≥ 15: moderate-to-severe AUD. ≥ 15: moderate-to-severe AUD. Korean version of the insomnia severity index (ISI-K) Insomnia severity was assessed using the Korean version of the Insomnia Severity Index (ISI-K), which is a validated adaptation of the ISI [27]. The ISI-K consists of seven items rated on a 5-point Likert scale, with total scores ranging 0–28. Higher scores indicate greater insomnia severity. Psychometric analysis of the ISI-K has shown strong reliability (Cronbach’s α = 0.72, item-total correlation 0.65–0.84) [28]. The severity of insomnia was classified as follows: < 8: no insomnia. < 8: no insomnia. 8–14: subthreshold insomnia. 8–14: subthreshold insomnia. 15–21: clinical insomnia (moderate severity). 15–21: clinical insomnia (moderate severity). 22–28: severe clinical insomnia. 22–28: severe clinical insomnia. A score of ≥ 15 was used as the clinical cutoff for insomnia. Korean occupational stress scale (KOSSⓇ19) Occupational stress was measured using the Korean Occupational Stress Scale (KOSS-19) [29]. This 19-item scale evaluates workplace stressors across eight subdomains: physical work environment, job demands, degree of autonomy, relationship conflicts, job instability, organizational system, lack of rewards, and occupational climate. Items were rated on a 4-point Likert scale, with subscale reliability coefficients (Cronbach’s α) ranging 0.73–0.94. Korean version of the connor–davidson resilience scale 2 (K-CD-RISC2) Resilience was assessed using the Korean version of the Connor–Davidson Resilience Scale (K-CD-RISC2), which evaluates psychological resilience over 1 month [30, 31]. The two-item scale is rated on a 5-point Likert scale (0 = not true at all, 4 = true nearly all of the time), with higher scores reflecting greater resilience. In this study, the K-CD-RISC2 demonstrated high reliability (Cronbach’s α = 0.93). Awareness and attitudes toward mental illness (AAM) Participants’ awareness of and attitudes toward mental illness were evaluated using the 2022 National Mental Health Literacy and Attitudes Survey [32]. This 21-item scale, rated on a 5-point Likert scale, evaluates three dimensions: understanding mental illness, negative perceptions of mental illness, and acceptance of mental illness. However, one item (AAM16) was not measured in the 2024 dataset. To ensure consistency between the development and temporal validation datasets, AAM16 was excluded from all analyses including model training and feature selection, and the remaining 20 items were utilized in both the 2023 development dataset and the 2024 temporal validation dataset. The subscale reliability coefficients (Cronbach’s α) were 0.94, 0.90, and 0.76, respectively. A full list of AAM scale items is provided in Table S2. Statistical approach overview To address the significant class imbalance in the original data, a balanced training set was created by undersampling the majority (non-PTSD) group. An XGBoost model was then trained to classify individuals with and without PTSD using features from mental health questionnaires, sleep surveys, and demographic data (Fig. 1). The contribution of each question to PTSD classification was assessed using the Shapley additive explanations (SHAP) method [33]. Based on the computed feature importance, the most predictive features were selected and new XGBoost models were developed using only these selected features. To ensure the models were optimized for real-world scenarios, these refined models were evaluated using a validation set that was specifically constructed to reflect the original imbalance ratio. This allowed us to select the best-performing model—one that maintained high accuracy with the minimum number of features under realistic conditions. After validation, the performance of this selected model was further confirmed on a testing set, which also preserved the original imbalanced distribution to provide an unbiased final assessment. Finally, the model’s generalizability was assessed using a temporal validation dataset collected in 2024 under the same survey framework. Feature importance for predicting the risk of PTSD. The importance of each feature is defined as the sum of the average absolute SHAP values in predicting PTSD. PHQ-9: Patient Health Questionnaire-9; K-CD-RISC: Korean version of the Connor-Davidson Resilience Scale; AAM: Awareness and Attitude Towards Mental Illness; KOSS: Korean Occupational Stress Scale; AUDIT: Alcohol Use Disorder Identification Test; ISI-K: Korean version of the Insomnia Severity Index; SHAP: Shapley Additive Explanations Model building To classify individuals, we used XGBoost, a tree-based machine-learning algorithm that employs gradient boosting [34]. This algorithm has been widely used in medical diagnostics because of its high accuracy, robust performance, interpretability, and ability to handle missing data [16, 17, 19, 20, 35–39]. Our dataset included a significantly larger number of non-PTSD individuals (n = 50,123) than PTSD individuals (n = 2,305). To prevent bias, a balanced training set was created. Specifically, the PTSD group was randomly partitioned into training (n = 1,613), validation (n = 461), and testing (n = 231) sets according to a 7:2:1 ratio. For the training set, an equal number of non-PTSD individuals (n = 1,613) were randomly selected from the available pool to match the PTSD group, ensuring a 1:1 ratio. In contrast, the validation and testing sets were constructed to reflect a realistic clinical scenario by maintaining the original 1:21 ratio of PTSD to non-PTSD individuals. This resulted in the inclusion of 9,681 non-PTSD individuals in the validation set and 4,851 in the testing set, both sampled from the remaining individuals not used in the training process. To optimize the XGBoost model, we tuned the hyperparameters using a stratified 5-fold cross-validation approach on the training set [40]. During this process, the training data were divided into five subsets. For each iteration, the model was trained on four subsets and validated on the remaining subset, using the F1-score as a performance metric. Bayesian optimization was employed to determine the hyperparameter combination that maximized the average F1-score across all five folds. The search space for optimization included max_depth (2–10), learning_rate (0.0001–0.1), n_estimators (50-1000), gamma (0.01–1), min_child_weight (1–10), max_delta_step (0–10), subsample (0.2–1), and colsample_bytree (0.1–0.99). After identifying the optimal hyperparameters, the final model was trained using the entire training dataset. For final training, the area under the receiver operating characteristic curve (AUROC) was used as the primary performance metric. Unlike the F1-score, which evaluates the performance at a specific classification threshold, the AUROC provides a comprehensive assessment of the model’s predictive ability across all possible thresholds. Feature selection To identify the most critical features contributing to PTSD classification, an XGBoost model was trained on the training set, and the feature importance was assessed on the validation set using the SHAP method [33]. SHAP is a game-theoretic approach that quantifies the impact of each feature on model predictions, providing the interpretability of classification results. In this analysis, positive SHAP values indicate an increased likelihood of PTSD classification, while negative values signify a contribution toward the non-PTSD group. The overall importance of each feature was determined by calculating the mean absolute SHAP values across the data points. The validation set was used for feature importance assessment and feature selection, while the test set was strictly held out and used only once for final performance evaluation to provide an unbiased estimate of generalization performance. Statistical analysis The classification model’s performance was evaluated using AUROC on the test dataset. The receiver operating characteristic (ROC) curve illustrates the trade-off between the false-positive rate (the proportion of non-PTSD individuals misclassified as having PTSD) and the true-positive rate (the proportion of correctly identified PTSD cases) across varying probability thresholds. The AUROC values ranged from 0 to 1, with higher values indicating superior discrimination between PTSD and non-PTSD individuals. XGBoost and other classification models were implemented using Python (version 3.7.16) with the Scikit–Learn [41] and XGBoost [34] libraries. Bayesian Optimization was performed using the Bayesian Optimization package [42]. Statistical analysis and machine-learning algorithms All statistical analyses and machine learning computations were conducted using SAS 9.4, R 4.3.1 (RStudio), and Python 3.7.16. Ethics approval The study protocol was reviewed and approved by the Institutional Review Board of Seoul National University Bundang Hospital (SNUBH). Informed consent was waived based on the following criteria: (1) research could not be conducted without the waiver; (2) the waiver did not adversely affect the rights and welfare of participants; and (3) data were anonymized before analysis, with all personal identifiers removed, ensuring minimal risk to participants. Results Data description The demographic and clinical characteristics of the datasets are presented in Table 1. The average age of the participants was 39.3 years (standard deviation [SD] = 9.4). When stratified by diagnosis, the mean age was 40.4 years (SD = 9.1) in the probable PTSD group and 39.3 years (SD = 9.4) in the non-PTSD group. The average work experience was 11.1 years (SD = 9.5), and the probable PTSD group had a slightly longer average work period of 12.0 years (SD = 9.4) compared to 11.1 years (SD = 9.5) than the non-PTSD group. Males comprised 90.2% of all firefighters, with a similar distribution in both the probable PTSD (90.3%) and non-PTSD (90.2%) groups. Table 1Demographic characteristics of participantsCharacteristicParticipants (n = 52,428)PTSD(n = 2,305)non-PTSD(n = 50,123)Male, n (%)47,308 (90.2%)2,081 (90.3%)45,227 (90.2%)Age (years), Mean (SD)39.3 (9.4)40.4 (9.1)39.3 (9.4)Work period (years), Mean (SD)11.1 (9.5)12.0 (9.4)11.1 (9.5)Marital status, n (%)Single18,130 (34.6%)739 (32.1%)17,391 (34.7%)Married33,438 (63.8%)1,486 (64.5%)31,952(63.7%)Divorced716 (1.4%)69 (3.0%)647 (1.3%)Separated36 (0.07%)2 (0.09%)34 (0.07%)Widowed108 (0.2%)9 (0.4%)99 (0.2%)Work type, n (%)12 h − 3 shift (21-day cycle)2,917 (5.6%)179 (7.8%)2,738 (5.5%)12 h − 2shift (9-day cycle)75 (0.1%)13 (0.6%)62 (0.1%)12 h − 2 shift (6-day cycle)57 (0.1%)10 (0.4%)47 (0.1%)12 h − 2 shift (3-day cycle)36,822 (70.2%)1,577 (68.4%)35,245 (70.3%)24 h − 2 shift (2-day cycle)96 (0.2%)5 (0.2%)91 (0.2%)12 h − 4 shift (4-day cycle)1,617 (3.1%)80 (3.5%)1,537 (3.1%)Day shift10,571 (20.2%)429 (18.6%)10,142 (20.2%)Others273 (0.1%)12 (0.5%)261 (0.1%) Demographic characteristics of participants For the temporal validation dataset collected in 2024 under the same survey framework, a total of 57,528 firefighters were included. The average age was 39.9 years (SD = 9.4), and males comprised 90.4% of the population. The prevalence of probable PTSD in the validation set was 5.0% (n = 2,848), showing comparable demographic features to the development dataset. Feature selection To identify the risk factors associated with probable PTSD without directly utilizing the PCL-5 checklist, we developed an algorithm capable of accurately detecting probable PTSD based on a reduced set of questions selected. These were selected from 92 questionnaire items across seven instruments (List of Traumatic Events, PHQ-9, AUDIT, ISI-K, KOSS-19, K-CD-RISC2, and AAM), along with 11 demographic characteristics. Initially, an XGBoost model was trained using all available input features to classify probable PTSD risk in the training set. After training, feature importance was assessed on the validation set using the SHAP method (Fig. 1), and then low-importance features were excluded to determine whether the model could maintain predictive accuracy with a smaller subset of features. To explore the optimal number of associative features, XGBoost models were constructed using varying numbers of top-ranked features, starting from 1 and increasing incrementally (e.g., 5, 10, …, 100). As the number of features increased, the AUROC for PTSD classification in the validation set improved (Fig. 2A). Notably, the AUROC plateaued after incorporating 10 or more features, with values consistently exceeding 0.9 (Fig. 2A; Table 2). To determine the smallest number of features required to achieve high AUROC values, we refined this approach by testing models with 1–15 features. The results indicated that models with seven or more features consistently achieved AUROC values of 0.9 or higher (Fig. 2A, star; Table 2). The sensitivity remained above 0.85, suggesting that a model with seven features provided sufficient accuracy for the classification of probable PTSD in firefighters (Fig. 2B; Table 2). Consequently, we developed a streamlined probable PTSD classification model, the Firefighter PTSD Latent Risk Evaluation (FLARE) (Fig. 3A). The top seven predictive features were decreased concentration (PHQ7), appetite changes (PHQ-5), feelings of worthlessness, excessive or inappropriate guilt (PHQ6), sleep disturbances (ISI4), perceived work pressure (KOSS5), poor adaptability to change (K-CD-RISC-1), and low perceived appropriateness of seeking professional help for close others (AAM6). Accuracy of XGBoost in predicting PTSD with varying numbers of feature. (A) The accuracy of PTSD prediction, as measured by AUROC, reaches 0.9 or higher, (dotted line) when seven or more features are used as input. (B) The sensitivity of PTSD prediction exceeds 0.85 (dotted line) when seven or more features are used as input Table 2XGBoost performance based on top-n important features n AUROCF1accuracyspecificitysensitivityAUPRC10.7993660.3638470.8917370.9017660.6811280.24039420.8266590.2587460.7994480.8008470.7700650.28732830.8333960.2499130.7881090.7886580.7765730.32961640.8452950.2791250.8212380.8240880.7613880.37435550.8549780.2787930.8209430.8237790.7613880.39930760.8911450.1849040.640110.6278280.8980480.40652270.9099050.2680280.7808120.7759530.8828630.42933880.9114620.2622530.7714450.7656230.8937090.43488590.900310.1302840.4076120.3805390.9761390.400453100.9063790.1448580.4761390.4523290.9761390.432901110.9155560.221930.7082430.6983780.9154010.458825120.9152770.2715720.7852490.7807040.8806940.454753130.9148730.2937980.8080260.8046690.8785250.449149140.915040.2730920.7858410.7811180.8850330.463977150.9166750.2809410.7920530.7872120.8937090.471736200.9026050.2249390.7186950.7101540.8980480.441971250.9048890.2272360.7196810.7107740.9067250.451398300.9034940.2261010.7246110.7169710.8850330.419592350.905540.2460440.7510350.7442410.8937090.460281400.9156040.2389940.7375270.729470.9067250.465054450.9125690.2409360.7440350.7369070.8937090.448418500.9168050.265430.7735160.7674830.9002170.467674550.9186870.2412320.7376260.7290570.917570.467531600.9251930.2891990.7988560.794030.9002170.499065650.9199670.2615190.7661210.7592190.9110630.474594700.9229810.2410870.7355550.7265780.9240780.478913750.9252710.2598040.7617830.7542610.919740.492046800.9200870.2915930.8030960.7988840.891540.416312850.9199610.2635760.7713470.765210.9002170.481014900.9225350.2579060.7616840.7545710.9110630.47172950.9268870.2721130.7768680.7701680.917570.5128461000.9256260.2662480.7695720.7624210.919740.487611030.9259310.3443190.8509170.8504290.8611710.510342 XGBoost performance based on top-n important features Accuracy of FLARE in predicting PTSD with a limited number of features. (A) Schematic representation of the FLARE. Diagnosing PTSD typically involves a combination of questionnaires, which can be time-consuming and complex. Our objective is to simplify this process by predicting the risk of PTSD. (B) The accuracy of PTSD prediction by the FLARE on the testing dataset, as measured by AUROC, exceeds 0.90 when using seven features as input. (C) The accuracy of PTSD prediction by the FLARE on temporal validation dataset, as measured by AUROC Results for model performance The performance of the FLARE model (Fig. 3A) was evaluated using a testing dataset. The FLARE model demonstrated high predictive accuracy, achieving an AUROC of 0.90 (Fig. 3B). By applying the optimal threshold derived from a previous study [43], based on the ROC curve, the model attained a sensitivity of 0.84 while minimizing false positives, resulting in a specificity of 0.82 (Fig. 3B, blue dots; Table 3). Table 3Performance of the FLARE model in test datasetsDatasetAUROCF1accuracyspecificitysensitivityAUPRCTest (2023)0.9048360.3037380.8239460.8223050.8441560.43947720240.9114160.3378770.8395560.8402160.8268960.464236 Performance of the FLARE model in test datasets To further validate the model’s generalizability, we applied FLARE to predict PTSD risk among firefighters using data from the 2024 survey, which was excluded from training and feature selection. The model maintained robust performance in this validation dataset, achieving an AUC of 0.91 (Fig. 3C). At the optimal threshold, the model yielded a sensitivity of 0.83 and a specificity of 0.84 (Table 3). These results indicated that the FLARE model is a reliable predictor of probable PTSD. Impact of each feature toward classification To further examine how FLARE differentiates individuals with probable PTSD from those without probable PTSD, SHAP values were computed for the seven features included in the model. A positive SHAP value indicates that a feature increases the likelihood of probable PTSD, whereas a negative value indicates the opposite. The distribution of the SHAP values was visualized using a beeswarm plot, where each dot represents an individual participant. The color gradient of the dots indicates the feature values, with blue denoting lower values and red representing higher values (Fig. 4). The plot reveals that individuals experiencing difficulty concentrating on tasks such as reading or watching, appetite changes, feelings of worthlessness or excessive/inappropriate guilt, sleep disturbances, perceived work pressure, poor adaptability to change, and low perceived appropriateness of seeking professional help for close others (red dots) were more likely to be classified as having probable PTSD than those who did not exhibit these characteristics (blue dots). Notably, the reversed trend observed for AAM6 (“Belief that family or friends should seek professional help for mental health problems”) suggests that individuals with lower scores on this item—reflecting a lack of recognition that mental illness is a treatable condition—were more likely to have probable PTSD. Overall, these findings indicate that severe depressive symptoms, a limited understanding of PTSD, and high occupational stress are strongly associated with an increased risk of probable PTSD. Impact of each feature on predicting PTSD. Beeswarm plot illustrating the influence of each feature on PTSD prediction in the testing set. Each point represents a single participant, with color indicating the feature value, ranging from blue (low values) to red (high values). The x-axis represents the direction of contribution: positive values indicate a contribution to increased PTSD risk Discussion We aimed to develop a machine-learning-based PTSD indicator (FLARE) to classify probable PTSD among South Korean firefighters using a concise set of clinically relevant psychosocial and demographic features. To our knowledge, this is the first study to develop a PTSD classification algorithm based on a broad range of clinical symptoms associated with related disorders among firefighters. The novel FLARE model consists of seven significant features: decreased concentration; appetite changes; feelings of worthlessness or excessive or inappropriate guilt; sleep disturbances; perceived occupational pressure; poor adaptability to change; and low perceived appropriateness of seeking professional help for close others. Previous studies have focused primarily on psychiatric diagnoses as risk factors for PTSD, often requiring extensive clinical expertise or lengthy questionnaires. In contrast, our approach relied on only seven easily answered items while achieving high classification accuracy (AUROC = 0.90) for probable PTSD in firefighters. This highlights the potential utility of the FLARE model in both clinical and operational settings, offering a practical tool for early identification of PTSD risk. The inclusion of depressive symptoms, such as feelings of worthlessness, excessive or inappropriate guilt, and concentration difficulties, as significant predictors align with prior findings that these symptoms are commonly associated with PTSD [24]. Among depressive symptoms, feelings of worthlessness, excessive/inappropriate guilt, and decreased concentration were linked to probable PTSD. Decreased concentration is a diagnostic criterion for both major depressive disorder and PTSD according to the DSM-5. Additionally, psychomotor agitation may be associated with hypervigilance, which is a key symptom of PTSD. Feelings of worthlessness, shame, and guilt overlap between depression and complex PTSD, with guilt being consistently linked to PTSD development and persistence [44, 45]. Guilt involves a negative evaluation of one’s own behavior, in which individuals perceive their actions or inactions as contributing to negative outcomes [46]. The severity of guilt following a traumatic event depends on the perceived level of personal involvement [47]. These feelings may contribute to trauma-related intrusion characteristics in patients with PTSD [48]. This underscores the importance of addressing guilt in PTSD interventions, to mitigate its role in the persistence of trauma-related intrusions. Additionally, this study highlighted the impact of occupational stress and mental health stigma on the risk of PTSD. High levels of stress, as reflected in perceptions of work pressure, emerged as a significant predictor, which is consistent with previous findings that occupational stress is a critical factor in mental health outcomes among first responders [11]. Firefighters experience substantial occupation-related stress due to physically demanding tasks, heavy workloads, frequent emergency calls, and irregular work hours [11]. One possible explanation is that increased perceived stress reduces access to the psychological resources necessary for coping with traumatic events, thereby increasing the likelihood of developing PTSD [49]. Interventions aimed at reducing occupational stress and strengthening the coping mechanisms of firefighters may help prevent PTSD in high-risk populations. Furthermore, a low perceived appropriateness of recommending professional help to family or friends — reflecting limited mental health literacy and stigma-related attitudes — was identified as a significant contributor to PTSD risk classification. This finding reflects the prevailing cultural attitudes in South Korea that discourage individuals from seeking psychological support [50, 51]. Addressing stigma and improving mental health literacy could be crucial steps in mitigating the risk of PTSD among firefighters. Several studies have identified stigma as a barrier to mental health treatment and a risk factor for various psychiatric disorders, including PTSD [15]. Stigma may limit individuals’ access to coping resources and strategies, while societal expectations of resilience among firefighters [52] may further discourage them from seeking professional services because of the fear of negative peer or administrative repercussions [13, 53]. Awareness-raising interventions should target not only high-risk individuals but also the broader firefighter community to enhance PTSD-related knowledge and reduce stigma. Psychoeducation emphasizing that PTSD is a medical condition requiring treatment rather than a sign of weakness may encourage affected individuals to seek help and reduce the incidence of PTSD. In our classification model, the type of trauma exposure exhibited low predictive power. However, other studies offer different perspectives. Pao et al. suggested that, while trauma exposure is a known risk factor, the type and severity of trauma play crucial roles in PTSD development among firefighters [54]. Their findings indicate that career firefighters who experience more severe and frequent traumatic events have higher PTSD rates than volunteer firefighters. A possible explanation for this discrepancy is that our study measured trauma exposure over the past year, rather than cumulative exposure over a firefighter’s career. Given that firefighters rotate through various assignments, their long-term exposure to trauma may be relatively uniform. Additionally, no demographic or occupational characteristics were among the top predictive features, suggesting that clinical factors are more influential than demographic factors in PTSD risk classification. Some features included in the model conceptually overlap with items from the PTSD Checklist for DSM-5 (PCL-5), such as sleep disturbances or concentration difficulties. However, the primary aim of the study was not to eliminate symptom similarity with existing PTSD instruments, but to assess whether PTSD risk could be detected without directly administering PTSD-specific questionnaires. In this context, overlapping features serve as indirect or proxy indicators of PTSD-related symptomatology that are commonly captured in routine health or occupational assessments. Importantly, the model did not incorporate the PCL-5 total score or diagnostic criteria, and feature importance analyses indicated that model performance was driven by the combined contribution of multiple features rather than by any single overlapping symptom. Notably, several influential features did not overlap with PCL-5 items, including variables related to perceived work pressure, adaptability, and stigma-related attitudes, which may reflect contextual psychosocial mechanisms influencing symptom expression and reporting behavior in occupational settings where mental health stigma is prevalent. This study has several limitations that must be addressed. The study population was predominantly male and exclusively Korean, limiting the generalizability of the findings to other populations. Additionally, the cross-sectional design precluded causal inferences regarding PTSD risk factors. Reliance on self-report questionnaires may also introduce bias; however, while self-report methods have inherent limitations, they remain practical and indispensable tools for large-scale screening, particularly when clinician-administered diagnostic interviews are logistically unfeasible. Nevertheless, this approach cannot fully address underreporting related to PTSD stigma or defensive responding, and some individuals with clinically significant symptoms may therefore not be captured by the reference labels used in this study. Accordingly, future studies should externally validate and calibrate the FLARE model using clinician-administered diagnostic interviews or other stronger reference standards in at least a subset of participants to further strengthen its diagnostic validity and generalizability. Future research should focus on longitudinal studies to better understand the causal relationships and temporal dynamics of risk factors for PTSD. Further validation of FLARE in diverse cultural and occupational contexts would enhance its generalizability and applicability. Moreover, the use of machine learning algorithms presents challenges for clinical implementation because these models often require specialized training, which makes them costly and time-consuming. In addition, the “black-box” nature of shortened questionnaires may hinder medical professionals’ understanding and trust. From a clinical perspective, FLARE may serve as a complementary screening and triage tool rather than a standalone diagnostic instrument. By leveraging psychosocial, occupational, and symptom-based features that are routinely collected in health or workplace assessments, FLARE could help identify individuals at elevated risk for PTSD who may not be flagged by conventional self-report screening alone, particularly in occupational settings where stigma or barriers to mental health disclosure limit the use of PTSD-specific questionnaires. In such contexts, FLARE may support clinical decision-making by prioritizing individuals for further evaluation, targeted psychoeducation, or referral for comprehensive diagnostic assessment without replacing clinician judgment. Recent studies have proposed automated methods that identify key questionnaire items and generate user-friendly score tables, allowing PTSD risk assessment through simple score summation [16, 17, 55]. These approaches provide a more transparent and interpretable alternative to machine-learning-based models while maintaining comparable accuracy. Incorporating such methods in future research could further refine the selection and scoring of key questionnaire items for PTSD classification. Conclusion In conclusion, this study developed and validated a machine learning algorithm that effectively classified probable PTSD risk among firefighters using a concise set of clinical and demographic features. The FLARE model provides a practical tool for early PTSD identification and can be seamlessly integrated into routine mental health screening. Given its high accuracy with only seven features, FLARE is an efficient and minimally burdensome method for assessing probable PTSD risk in firefighters. Its implementation could facilitate early intervention, prioritize individuals in need of further psychological evaluation and treatment, and ultimately improve mental health outcomes and overall well-being. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Acknowledgements We appreciate the dedication of Young Kuk Ju, MuKi Kim, and Young HakShin to the Fire Policy Division of the Health Management Team of the National Fire Agency. Abbreviations Post-traumatic stress disorder Alcohol use disorder PTSD Checklist for DSM-5 Patient Health Questionnaire-9 Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition Alcohol use disorders identification test World Health Organization Korean version of the Insomnia Severity Index Korean Occupational Stress Scale Korean version of the Connor–Davidson Resilience Scale Awareness and attitudes toward mental illness Shapley additive explanations Area under the receiver operating characteristic curve Standard deviation Firefighter PTSD Latent Risk Evaluation Author contributions Seok Joo Chae: Conceptualization, Data curation, Formal analysis, Methodology, Writing – original draft. Johanna Inhyang Kim: Conceptualization, Data curation, Formal analysis, Methodology, Writing – original draft. Hyeontae Jo: Formal analysis, Methodology. Yun Min Song: Formal analysis, Methodology. Saebom Jeon: Formal analysis, Methodology. Sun-Young Moon: Conceptualization, Data curation, Investigation, Resources. Ji-Hye Lee: Conceptualization, Data curation, Investigation, Resources. Sohee Oh: Formal analysis, Methodology. Seockhoon Chung: Conceptualization, Methodology. Jae Kyoung Kim: Conceptualization, Methodology, Supervision, Writing – review & editing. Jeong-Hyun Kim: Conceptualization, Funding acquisition, Supervision, Writing – review & editing. Funding This research was partially supported by grants from the National Fire Agency of Korea, Institute for Basic Science (IBS-R029-C3), National Research Foundation of Korea (NRF) (RS-2024-00357912, RS-2022-nR068758), and Ajou University Grant (S-2026-G0001-00017). Data availability Access to data is restricted. Declarations Ethics approval and consent to participate The study protocol was reviewed and approved by the Institutional Review Board of Seoul National University Bundang Hospital (SNUBH). Informed consent was waived based on the following criteria: (1) research could not be conducted without the waiver; (2) the waiver did not adversely affect the rights and welfare of participants; and (3) data were anonymized before analysis, with all personal identifiers removed, ensuring minimal risk to participants. The study was conducted in accordance with the principles for the protection of the rights of study participants as outlined in the Declaration of Helsinki. Consent for publication Not applicable. Disclosures The authors have no actual or potential competing financial interests to declare. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Seok Joo Chae and Johanna Inhyang Kim first authors contributed equally to this manuscript. Jae Kyoung Kim and Jeong-Hyun Kim corresponding authors contributed equally to this manuscript. Contributor Information Jae Kyoung Kim, Email: jaekkim@kaist.ac.kr. Jeong-Hyun Kim, Email: jhkim@snubh.org. References Associated Data Data Availability Statement Access to data is restricted.