Predicting Probable Persistent PTSD Following the Sewol Ferry Disaster: Development of an AI Algorithm Based on Psychological Assessments and Biological Markers
Article information
Abstract
Objective
Prolonged post-traumatic stress disorder (PTSD) carries substantial personal and public-health costs, yet early identification of individuals at risk remains difficult. This study aimed to develop a novel artificial intelligence (AI)-based predictive algorithm using psychological, biological, and psychosocial data collected at initial assessment to enhance early identification of individuals at heightened risk for probable persistent PTSD.
Methods
This study included 88 bereaved family members of the victims of the Sewol ferry disaster, divided into probable persistent PTSD (n=67) and probable remitted (n=21) groups based on 4-year follow-up assessments. Demographic, blood test, and psychological data were collected during initial evaluations. Models compared linear discriminant analysis (LDA) with tree-based learners under stratified cross-validation; class imbalance was addressed with Borderline Synthetic Minority Oversampling Technique, and recursive feature elimination identified parsimonious predictors.
Results
The LDA model demonstrated the highest performance, with an area under the receiver operating characteristic curve of 0.858. Psychosocial features dominated prediction: higher anxiety, depression, insomnia, and intrusive rumination increased risk, whereas greater positive resources and functional social support were protective. Routine physiological markers showed limited incremental value.
Conclusion
Findings support a practical intake pathway in which brief psychosocial measures are used with an interpretable classifier to triage high-risk individuals to targeted interventions. External validation, calibration, decision-curve analysis, and broader biomarker panels are needed to confirm transportability and optimize clinical utility.
INTRODUCTION
Post-traumatic stress disorder (PTSD) impairs daily functioning, relationships, and occupational performance, and its chronicity is linked to elevated risks for suicide and diverse medical comorbidities [1]. Although symptoms often diminish over time, a substantial subset, approximately 39% in prior work, develops a chronic course [2]. Tools capable of estimating an individual’s likelihood of persistent PTSD at baseline are therefore essential for preventive care and personalization of treatment plans [3,4]. Furthermore, bereaved families of individuals who died because of traumatic events frequently experi-ence PTSD symptoms with heterogeneous trajectories [5]. The Sewol ferry disaster in 2014 profoundly affected Korean society, with bereaved families reporting sustained psychological distress [6]. Notably, the Sewol incident involved high school students who were on a school trip when the ferry tragically sank, amplifying the emotional impact on families and the broader community. The unique nature of the Sewol incident, involving young passengers and controversial rescue efforts, has intensified the emotional burden on the bereaved families [7]. Given the scale and characteristics of this event, identifying individuals at initial assessment who are likely to follow a prolonged course of PTSD is a clinical priority to enable timely and targeted support. Accordingly, long-term monitoring and early risk stratification are crucial for effective intervention and treatment delivery among bereaved families affected by large-scale traumatic loss.
Prior meta-analyses and prospective cohort studies have identified risk factors for PTSD persistence across three temporal domains [8,9]. Pre-trauma vulnerabilities include family psychiatric history, personal history of mood or anxiety disorders, and early-life adversity such as childhood trauma. Peritrauma factors encompass trauma severity, perceived life threat, and peritraumatic dissociation, all of which have shown robust associations with chronic PTSD trajectories. Post-trauma indicators— including initial PTSD symptom severity, comorbid depression and anxiety, maladaptive cognitive processes such as rumination and avoidance, sleep disturbance, ongoing life stressors, and lack of social support—have emerged as particularly strong predictors of non-recovery [10,11]. The present study focuses primarily on post-trauma psychosocial indicators measurable at baseline assessment, as these are most amenable to scalable screening in disaster mental health settings. However, the incremental predictive value of routine physiological indices, such as lipid profile, inflammatory, glycemic, thyroid, and hepatic markers, remains uncertain [9,12]. In this context, if robust prognostication can be achieved using blood tests available from routine health-screening panels, clinical utility and scalability would be heightened in disaster mental-health settings by enabling low-burden risk stratification at the time of initial contact. Furthermore, machine learning (ML) framework can generate calibrated probability estimates and support decision-curve analysis, making referral thresholds explicit for triage. Cross-validated, feature-selected models are also well suited to small samples with correlated predictors and imbalanced outcomes, and, when implemented with interpretable classifiers, can be embedded into routine workflows without overburdening clinicians [13]. However, this review focused on early post-trauma symptom assessment rather than long-term prognostication. Likewise, a ML classifier that separated remission versus non-remission using 6-month symptom status achieved approximately 0.70 area under the receiver operating characteristic curve (AUC) under external validation, indicating promise for short-horizon prediction [14]. Because those models used outcomes defined by later symptoms within the first 6 months, the clinical utility of artificial intelligence (AI) for long-term prognosis remains insufficiently established. Related work employing natural language models to predict 2-year PTSD persistence has also been proposed, however, a 2-year horizon may still fall within the window of symptom fluctuation and may not capture durable chronicity, further underscoring the need for longer-horizon prognostic studies [15].
This study aims to develop and evaluate a ML-based predictive algorithm using baseline demographic, physiological, and psychosocial data to identify bereaved family members at risk for 4 years probable persistent PTSD symptoms after the Sewol ferry disaster. Higher baseline anxiety, depressive symptoms, insomnia, and intrusive rumination are hypothesized to predict PTSD persistence, whereas greater functional social support and positive psychological resources are expected to predict symptom relief. A ML approach integrating these domains is further posited to outperform models relying on any single domain, with psychosocial features contributing more strongly to prediction than routine physiological indices. If the algorithm achieves accuracy above 80%, leveraging the most informative baseline indicators would enable AIassisted early risk stratification and support proactive, personalized interventions for individuals most vulnerable to chronic PTSD.
METHODS
Participants
Recruitment began in 2015, roughly 1 year after the disaster, following briefings with family representatives and public announcements through the Ansan Mental Health Trauma Center. With the center’s support, 254 bereaved families enrolled in total. From enrolment onward, every participant completed an annual mental health assessment, and fasting blood tests were obtained every 2 years as part of a routine health screening protocol. Because enrolment occurred in waves across years, earlier waves accumulated more follow up assessments than later ones. For example, those who entered in 2015 could contribute as many as seven annual assessments by 2021, whereas those who entered in 2020 could contribute only two by that same calendar end point. This staggered structure explains why eligibility for 4 year outcome ascertainment rests chiefly with the earlier waves, while later waves are still accruing follow up time [16]. For the present analyses, data from 287 participants observed between 2015 and 2021 were reviewed. Of these, 165 were excluded due to incomplete baseline clinical assessments, missing baseline blood tests, or unavailable clinical assessment data at the 4-year follow-up, leaving 122 participants with complete baseline and follow-up data. An additional 34 participants were excluded because their baseline PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (PCL-5) scores were below the clinical threshold (<33), indicating subthreshold PTSD symptoms at initial evaluation. Because this study aimed to predict persistence versus remission among individuals with probable PTSD at baseline, those without clinically significant initial symptoms were not included in the target population. The final analytic sample comprised 88 participants (Figure 1).
Participant selection process for baseline and 4-year follow-up PTSD analysis. PTSD, post-traumatic stress disorder; PCL-5, PTSD Checklist for Diagnostic and Statistical Manual of Mental Disorders, 5th Edition.
Exclusion criteria for this analysis included patients who missed two or more questions in the initial psychosocial questionnaire, those who did not complete the PTSD scale 4 years later, and among those who completed the initial PTSD scale but did not undergo a health examination. All participants willingly provided written informed consent, recognizing their right to withdraw from the study at any point. Furthermore, this study protocol was approved following a review by the Institutional Review Board of Seoul St. Mary’s Hospital, The Catholic University of Korea (approval no. KC15OIMI0261), and the National Medical Center (approval no. H-1505-054-002). All participants provided informed consent for their data to be used for future research purposes, and the application of AI techniques to the dataset was conducted under the scope of the original IRB approval.
Demographic and physiological data
Baseline assessments were conducted at the time of enrollment, which occurred in annual waves beginning in 2015, approximately 12 months after the Sewol ferry disaster (April 16, 2014). The outcome was defined as PTSD symptom status at 4 years after the baseline evaluation, assessed using the PCL-5. Fasting blood samples and psychological questionnaires were collected on the same day during the baseline visit to ensure temporal alignment of physiological and psychosocial data. Demographic data were collected, including sex, age, and body mass index (BMI). Additionally, systolic and diastolic blood pressure (BP) were measured, and their mean values were calculated. These physiological parameters provided baseline characteristics essential for understanding the participants’ overall health status. Various blood tests were conducted to evaluate comprehensive aspects of physical health. Liver function was assessed by measuring serum glutamicoxaloacetic transaminase and serum glutamic-pyruvic transaminase [17] levels. Nutritional status was evaluated through protein and albumin concentrations. Renal function was determined by measuring blood urea nitrogen [18] and creatinine levels. Inflammatory markers were assessed using C-reactive protein levels. Glycemic control was evaluated with hemoglobin A1c tests, while lipid profiles were determined by measuring low-density lipoprotein, triglycerides, and high-density lipoprotein (HDL). Thyroid function was assessed by measuring thyroid-stimulating hormone and free thyroxine levels. These laboratory parameters provided a comprehensive overview of the participants’ physical health, which was crucial for analyzing potential biological markers associated with probable persistent PTSD.
Selection of physiological indices followed two principles. First, all measures are readily available in routine clinical practice, which supports feasibility and scalability for intake workflows. Second, the panels capture domains that have been variably linked to stress related pathophysiology and long term symptom persistence, including inflammation, metabolic and endocrine regulation [19-21]. In keeping with the study aim, these indices were included to test their incremental prognostic value over concise psychosocial measures while maintaining a pragmatic pathway for real world deployment.
Clinical assessments
The primary tool for assessing PTSD symptoms was the PCL-5 validated in Korean [22]. The PCL-5 is a 20-item self-report measure that aligns with the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition criteria for PTSD, offering a reliable and standardized assessment of symptom severity [23,24]. Each scored on a scale of 0 to 4, resulting in a total score range of 0 to 80. A cutoff score of ≥33 was used to de-fine probable PTSD, based on validation studies demonstrating optimal diagnostic efficiency at this threshold [25]. Although alternative cutoffs have been proposed in the literature (e.g., 31 for higher sensitivity or 38 for higher specificity), the threshold of 33 is widely adopted in research settings and was selected to balance sensitivity and specificity. At the 4-year followup, the PCL-5 was re-administered to evaluate the persistence of PTSD symptoms, and this follow-up status served as the study outcome, with probable persistent PTSD defined as a PCL-5 score of 33 or higher and probable remitted as a score below 33. This operationalization is grounded in longitudinal trajectory research showing that earthquake survivors with clinically significant symptoms at approximately 52 months tend to follow worsening or chronic courses, and in oncology cohorts where PTSD persisting for up to 4 years has been interpreted as chronic; taken together, these findings support the use of a 48-month anchor as a reasonable index of chronicity [26,27]. It should be noted that the PCL-5 is a self-report screening instrument rather than a clinician-administered diagnostic interview, and the current design did not include intermediate assessments between baseline and the 4-year follow-up. Consequently, the outcome classification reflects symptom status at a single time point and may not capture intermittent remission/relapse patterns. Accordingly, the terms “probable persistent PTSD” and “probable remitted” are used throughout this manuscript to acknowledge these measurement limitations.
Depressive symptoms were measured with the validated Korean version of Patient Health Questionnaire-9 (PHQ) with 9 items scored from 0 to 3; conventional thresholds indicate mild symptoms from 5 to 9, moderate from 10 to 19, and severe at 20 or higher [28,29]. Anxiety levels were measured with validated Korean version of the Generalized Anxiety Disorder-7 (GAD), comprising 7 items scored from 0 to 3, with total points of mild from 5 to 9, moderate from 10 to 14, and severe anxiety from 15 to 21 [30,31]. Insomnia was measured with the validated Korean Insomnia Severity Index (ISI) (7 items scored 0–4; total 0–28) with clinically significant insomnia defined as a score of at least 16 [32,33].
To evaluate effect of trauma and psychological factors, the validated Korean version of post-traumatic growth inventory measured positive psychological changes, consisting of 10 items (0–5 scale, total score up to 50), with higher scores reflecting greater positive changes [34,35]. Rumination was measured with the Korean version of Event-Related Rumination Inventory (ERR) (20 items; intrusive and intentional subscales; total 0–60), which separately evaluated negative intrusive rumination (ERR-1) and potentially adaptive intentional rumination (ERR-2) [24,36,37]. Coping was measured with the Korean Brief COPE (28 items) and summarized into problem-focused, emotion-focused, and dysfunctional or avoidant domains [38]. Functional social support was measured with the functional social support questionnaire (FSSQ) (total 14–70, higher scores indicate greater support) and Korean version was validated [39,40]. Coping was assessed with the Brief COPE and summarized into three domains: problem-focused strategies that actively address or modify the stressor, emotion-focused strategies that regulate affective responses, and avoidant strategies that minimize or evade the stressor; this taxonomy clarifies how coping styles relate to PTSD trajectories [41]. Social support was measured with the FSSQ (14 items scored 1–5; total 14–70), with higher scores indicating greater support [42]. Positive psychological resources—including optimism, purpose or hope, selfcontrol, social support, and care—were measured with the Positive Resources Test (POREST) (23 items scored 1–5; total 23–115), where higher scores indicate greater resources [43].
The comprehensive nature of these assessments allowed for a multifaceted analysis of the factors contributing to probable persistent PTSD, facilitating the identification of key psychological predictors and enhancing the development of the AI algorithm for PTSD prediction.
Statistical analysis
Statistical analyses were performed to examine the relationships between demographic, physiological, laboratory, and psychological variables with probable persistent PTSD outcomes. Categorical variables, such as gender, were analyzed using the chi-square test to assess differences in distribution between groups. For continuous variables that met the assumptions of normality, including the ISI and POREST scores, independent samples t-tests were conducted to compare means between participants with and without probable persistent PTSD. For continuous variables that did not follow a normal distribution, non-parametric Mann–Whitney U tests were utilized to evaluate differences between groups. This approach ensured the appropriate handling of data based on their distribution characteristics, enhancing the validity of the statistical inferences. All statistical analyses were performed using SPSS version 21 (IBM Corp.). A significant level of p<0.05 was set for all tests.
AI analysis for predicting probable persistent PTSD
To develop and validate an AI algorithm capable of predicting probable persistent PTSD, a structured and methodical approach was implemented, addressing the inherent class imbalance in the dataset and optimizing model performance through advanced techniques. The dataset, consisting of individuals with probable persistent PTSD and probable remitted, was divided using stratified 10-fold cross-validation [44]. This method ensured that the distribution of prolonged and probable remitted cases remained balanced across both training and test sets, minimizing potential biases during model training and evaluation. Given the imbalanced nature of the dataset, particularly the underrepresentation of the probable remitted group, the Borderline Synthetic Minority Oversampling Technique (SMOTE) was employed [4]5. Unlike traditional oversampling methods, Borderline SMOTE focuses on generating synthetic samples near the decision boundary where the minority class (probable remitted) is most vulnerable to misclassification [46]. This approach enhanced the representation of probable remitted cases in critical regions of the feature space, improving the model’s capacity to generalize to underrepresented cases.
To classify probable persistent PTSD and probable remitted, two categories of ML models were explored. linear discriminant analysis (LDA) and logistic regression (LR) models were utilized for their interpretability and computational efficiency, serving as baseline classifiers to benchmark performance against more complex methods [47]. Advanced tree-based models, such as Extreme Gradient Boosting (XGB) and Light Gradient Boosting Machine (LGBM), were employed due to their ability to handle complex, non-linear relationships within the data and their robust performance in predictive modeling tasks [48]. To ensure optimal performance of both linear- and tree-based models, Bayesian optimization was used for hyperparameter tuning. Bayesian optimization systematically searched the hyperparameter space to identify configurations that maximized model performance while minimizing computational costs [49]. This approach enabled efficient and effective fine-tuning of key model parameters, ensuring the best possible predictive accuracy.
Features selection for AI model
To optimize the predictive performance of the AI models, recursive feature elimination (RFE) was applied to identify the most important features among the 22 clinical and psychological variables collected. RFE is a widely used ML technique that iteratively evaluates feature importance based on the contribution of each variable to the model’s predictive performance [50,51]. In each iteration, the feature with the least importance is removed, and the process is repeated until the desired number of features is achieved [52]. The RFE process was conducted using LR as the base model, which is well-suited for small datasets and binary classification tasks such as distinguishing probable persistent PTSD from probable remitted. The features were ranked based on their contribution to the model, with those demonstrating the highest predictive power prioritized. To enhance the model’s predictive accuracy and efficiency, only the Rank 1 features were selected for further analysis. This selection was based on the principle of balancing model complexity and performance, particularly given the constraints of a small dataset [53]. For final model optimization, a greedy approach was employed, where the top-ranked features were incrementally added to assess their impact on predictive performance. This iterative process ensured that the selected feature set maximized the model’s ability to classify probable persistent PTSD and probable remitted accurately, without introducing unnecessary complexity or overfitting [54].
This iterative process ensured that the selected feature set maximized the model’s ability to classify probable persistent PTSD and probable remitted accurately, without introducing unnecessary complexity or overfitting [51]. To assess the stability of the feature selection process, we conducted 1,000 bootstrap iterations and calculated the selection frequency for each variable. Furthermore, while SHAP values were used for multimodel comparison in the main text, the standardized discriminant coefficients for the final LDA model were calculated and provided in Supplementary Figure 1 to ensure traditional interpretability.
RESULTS
Demographics and physiological features
Among the 287 participants included in the initial evaluation, 165 participants were excluded owing to incomplete data. Among the remaining 122 participants, 88 participants with significant PTSD symptoms at the initial assessment who had completed the PTSD symptom assessments at the 4-year follow-up examination were included in the final analysis. Sixtyseven (76.1%) and 21 (23.9%) of these 88 participants were classified as having probable persistent PTSD and probable remitted, respectively.
No statistically significant differences were observed between the probable persistent PTSD and probable remitted groups in terms of demographic characteristics such as sex, age, mean BP, and BMI. Similarly, no statistically significant differences were observed between the groups in terms of the blood test results, such as lipid profiles, glycemic control markers, inflammatory markers, liver function, thyroid function, and nutritional status (Table 1).
Clinical demographics and blood features between probable persistent PTSD and probable remitted
Clinical assessments
The levels of generalized anxiety symptoms, as measured by the GAD scale, in the participants with probable persistent PTSD were significantly higher than that in those with probable remitted (13.78±5.04 vs. 7.76±4.46, p<0.001). Similarly, the severity of the depressive symptoms, as assessed by PHQ in the probable persistent PTSD group (21.72±4.24) was significantly higher than that in the probable remitted group (17.38±3.37, p<0.001).
The ERR scores for intrusive rumination in the probable persistent PTSD group were significantly higher than those in the probable remitted group (23.73±5.56 vs. 19.57±7.03, p=0.012). In contrast, emotional coping, as measured by the emotional subscale of Brief COPE (COPE_2), in the probable remitted group (20.29±5.02) was significantly higher than that in the probable persistent PTSD group (17.66±4.99, p=0.011). Furthermore, positive resources, as assessed by POREST, in the probable remitted group (66.57±14.07) were significantly greater than that in the probable persistent PTSD group (57.25±15.33, p=0.015). These findings suggest that compared with individuals with probable persistent PTSD, those with probable remitted possess more positive psychological resources and adaptive coping mechanisms (Table 2).
Psychiatric characteristics between probable persistent PTSD and probable remitted
Classification of probable persistent PTSD and probable remitted by AI algorithms
LDA achieved the highest performance among the models, with an AUC of 0.858±0.110 indicating excellent discriminatory ability. Furthermore, LDA also demonstrated the highest accuracy (0.729±0.110), precision (0.908±0.100), and specificity (0.700±0.332), indicating its robustness in distinguishing probable persistent PTSD from probable remitted. LR also exhibited good performance, with an AUC of 0.823± 0.173, accuracy of 0.762±0.091, and specificity of 0.700±0.332.
Compared with the linear models, the tree-based models XGB and LGBM exhibited relatively lower performances. XGB achieved AUC and accuracy of 0.776±0.128 and 0.750 ±0.100, respectively. LGBM achieved AUC and accuracy of 0.782±0.133 and 0.728±0.125, respectively. The recall scores of both tree-based models were higher (XGB: 0.821±0.130; LGBM: 0.788±0.173), indicating their sensitivity in detecting probable persistent PTSD but at the cost of lower specificity (Table 3). These findings were further corroborated by bootstrap-based optimism correction with 1,000 iterations, which confirmed the superior discriminatory performance of LDA (AUC=0.803 [95% confidence interval, CI: 0.635–0.929]) compared with the tree-based models (XGB: AUC=0.741 [95% CI: 0.574–0.893]) (Supplementary Table 1).
Artificial intelligence algorithm functions (mean [95% CI])
The receiver operating characteristic curve comparison depicted in Figure 2 further illustrates the superior AUC of LDA compared with those of the other models, highlighting its reliability as the optimal model for predicting probable persistent PTSD in the present study. The recall of the treebased models was slightly higher; however, the trade-offs in precision and specificity suggest limitations in the overall predictive balance (Figure 2).
ROC curve comparison between artificial intelligence models. ROC, receiver operating characteristic.
Key features of AI algorithms
To assess the robustness of the feature selection process, we evaluated the selection frequency of each variable using 1,000 bootstrap iterations. The core psychosocial predictors (GAD, PHQ, and ERR-1) demonstrated exceptional stability, being selected in over 99% of the iterations. Detailed selection frequencies for all features are provided in Supplementary Figure 2. GAD emerged as the most influential feature in both models, indicating that it was the most important feature facilitating the differentiation of probable persistent PTSD from probable remitted. HDL and PHQ were consistently ranked as the second- and third-most important features, respectively, emphasizing the interplay between mental health symptoms and physiological markers. ERR-1 (negative intrusive rumination), FSSQ, and POREST were the other notable features that collectively reflected the critical roles of cognitive, emotional, and social factors in PTSD outcomes. The contribution of the ISI to the models was less significant; nevertheless, it played a meaningful role in the classification. The alignment of feature rankings between LDA and LR reinforced the robustness of the selected features. These findings provide strong evidence regarding their relevance in predicting probable persistent PTSD (Figure 3).
Key features of linear-based models. A: Feature importance (linear discriminant analysis). B: Feature importance (logistic regression). GAD, Generalized Anxiety Disorder-7; HDL, high-density lipoprotein; PHQ, Patient Health Questionnaire-9; ERR-1, Event-Related Rumination Inventory negative intrusive rumitation; FSSQ, functional social support questionnaire; ISI, Insomnia Severity Index; POREST, Positive Resources Test.
The PHQ was consistently ranked the most important feature in both tree-based models, underscoring the significant impact of depressive symptoms on PTSD outcomes. Among physiological indices, HDL was the only marker that contrib-uted meaningfully and did so with modest but consistent influence. GAD were also influential features, indicating their importance across the model types. ERR-1 was also identified as a prominent feature in the LGBM and XGB models, indicating the critical role of intrusive rumination in probable persistent PTSD. POREST, which represents positive psychological resources, was ranked high in both models, further emphasizing their protective role in recovery from PTSD. Notably, although the FSSQ and ISI were ranked lower, they made a meaningful contribution to the predictive performance of the models (Figure 4).
Key features of tree-based models. A: Feature importance (XGB classifier). B: Feature importance (LGBM classifier). PHQ, Patient Health Questionnaire-9; HDL, high-density lipoprotein; GAD, Generalized Anxiety Disorder-7; ERR-1, Event-Related Rumination Inventory negative intrusive rumitation; POREST, Positive Resources Test; FSSQ, functional social support questionnaire; ISI, Insomnia Severity Index.
DISCUSSION
Based on this study, probable persistent PTSD was present in 76% of the cohort at 4 years (67 of 88), and demographic characteristics and routine physiological markers, including standard blood tests, did not differ between prolonged and probable remitted groups. Additionally, baseline psychosocial profiles were strongly discriminative. Probable persistent PTSD was associated with higher depression, higher anxiety, and greater intrusive rumination, whereas relief was linked to greater emotional coping and higher positive resources. Among physiological indices, HDL was the only marker that showed a meaningful association with long term prognosis. Among the AI models, an interpretable linear classifier using LDA provided the best overall discrimination with a strong area under the curve. Clinically, these findings indicate that intaketime screening should prioritize brief psychosocial measures over routine laboratory panels, and that a lightweight, interpretable classifier can support early, low-burden risk stratification to direct stepped-care interventions toward modifiable targets such as insomnia, rumination, anxiety, and depression while actively bolstering social support and positive psychological resources.
First, in this cohort, probable persistent PTSD persisted in 76 percent of participants over approximately 4 years, a proportion higher than several prior reports that used different populations, instruments, and follow-up intervals. Research has shown that 4.8% of individuals experienced PTSD 12–15 years after a suspected serious injury, highlighting that PTSD can persist far beyond 4 years [55]. Globally, about half of PTSD cases are reported to be persistent, with the 12-month prevalence among lifetime cases estimated at 8.8% (95% CI, 5.5%– 13.5%), indicating the long-lasting nature of the condition for many affected individuals [18,56]. The findings from the 2004 Indian Ocean disaster study indicates that approximately 40% of cases with PTSD persisted for longer than 36 months [57]. Differences across studies likely reflect methodological factors, including case definitions, baseline symptom severity, bereavement status, enrollment criteria, and timing of follow-up, as well as our sample’s enrichment for individuals with marked initial PTSD symptoms. Because the present data do not directly assess cultural or situational determinants, no inferences about those influences are drawn here. Instead, the findings highlight the need for long-term monitoring and care for bereaved individuals at high risk of symptom persistence and underscore the importance of external validation in broader, population-based samples.
Secondly, the study demonstrated no statistically significant differences in demographic and physiological features, including blood test markers, aligning with prior research that underscores the limitations of traditional markers in differentiating PTSD outcomes. Previous research on the Sewol ferry disaster found that participants reporting clinically significant PTSD symptoms exhibited lower serum HDL levels than those without PTSD symptoms [58]. However, no differences of lipid profiles were observed between probable persistent PTSD and probable remitted groups. It is worth noting that, despite this lack of univariate significance, HDL contributed to the multivariate ML models. This apparent discrepancy reflects the distinction between univariate and multivariate analysis: a variable may provide incremental predictive value in the context of other features even when it does not independently differentiate groups. HDL may capture physiological variance that is complementary to the psychosocial predictors, thereby contributing to overall model performance. This suggests that the biological mechanisms associated with the onset of PTSD symptoms may differ from those related to the persistence of symptoms. Further previous studies highlighted the reliance of symptom-based diagnostics on subjective reporting and their weak correlation with objective cognitive or biological measures [59]. Together with the heterogeneity of PTSD trajectories reported in the literature, these findings support prioritizing psychosocial assessment for prognostication and treating routine physiology as exploratory in models of long-term course, with future studies needed to test more specific neuroendocrine or genetic markers and to externally validate these null physiological contrasts for persistence [17].
There were significant differences in psychiatric assessments, with probable persistent PTSD linked to higher levels of depression, anxiety and intrusive rumination. This finding aligns with previous results, indicating that intrusive rumination was a risk factor for PTSD even among the bereaved families affected by the Sewol ferry disaster [60]. Studies have also found that patients with probable persistent PTSD exhibit higher levels of rumination compared to those who experience symptom relief, with intrusive rumination strongly linked to PTSD severity and the maintenance of symptoms [61-63]. It has also been shown that ruminating about trauma often leads to intrusions, unproductive thoughts, and difficulties in emotion regulation, contributing to higher levels of anxiety and depression and lower emotional coping in probable persistent PTSD cases. In contrast, emotional coping and positive resources were significantly higher in the probable remitted group. Emotional coping has been identified as key factors associated with PTSD symptom relief [64]. Additionally, in this study, individual positive resources, including optimism, purpose/hope, self-control, social support, and care, emerged as critical factors distinguishing probable persistent PTSD from probable remitted. These findings accord with evidence that cultivating gratitude and other positive emotions reduces distress and enhances well-being in PTSD [65,66]. More broadly, strengthening positive affect, adaptive thinking, and emotional coping skills appears central to resilience, helping individuals confront and integrate distressing memories and thereby supporting recovery.
Additionally, this study is particularly noteworthy for its attempt to predict the persistence of PTSD using AI algorithms based on blood tests and psychometric evaluations conducted at the first visit. Currently, no AI-based algorithms specifically designed to predict PTSD persisting for more than 4 years have been published. The algorithm developed in this study demonstrates significant clinical relevance by achieving substantially improved accuracy using advanced techniques such as SMOTE for addressing class imbalance and RFE for feature selection. The most informative predictors were psychosocial measures (GAD, PHQ, ERR), with additional contributions from social support and positive resources; routine laboratory markers added limited signal. Clinically, this yields a lowburden intake tool for early risk stratification that can direct tailored interventions toward modifiable targets such as rumination, anxiety, and depression while concurrently bolstering coping and resilience. Also, the LDA classifier achieved the strongest overall discrimination (AUC≈0.86) with favorable accuracy, precision, and specificity; tree-based models showed higher recall at the cost of lower specificity. This pattern supports use of a lightweight, interpretable model for intake-time triage.
This study developed an AI algorithm using long term follow up data to identify probable persistent PTSD early and demonstrated high predictive performance. Across candidate approaches, the findings indicate that interpretable models are best suited for long horizon prognosis from an initial evaluation. An interpretable linear classifier achieved robust discrimination and is appropriate for low burden intake screening to identify individuals at elevated risk of persistent symptoms. These results translate into a practical intake pathway in which clinicians administer brief psychosocial measures at first contact, compute a calibrated probability of 4 year persistence with the interpretable classifier, and then triage high risk individuals to targeted interventions. Priority targets include rumination focused cognitive strategies, cognitive behavioral therapy for insomnia when indicated, and first line treatments for anxiety and depression, together with structured enhancement of functional social support and positive psychological resources. Because routine laboratory panels contributed little incremental signal beyond psychosocial profiles, they are not recommended as standalone prognostic tools, although they may be considered adjunctive to psychosocial risk assessment within stepped care workflows.
Despite these strengths, this study has several limitations. First, the study focused on PTSD symptoms but did not assess prolonged grief disorders (PGD). PTSD and PGD are distinct but potentially co-occurring syndromes following sudden loss, making them significant differences. Rumination, one of the strongest predictors here, is implicated in the maintenance of both conditions, but the shared cognitive pathway could not be tested without concurrent PGD measurement. Future work should include structured clinical interviews, broader biomarker panels, external validation with decisioncurve analysis, and parallel assessment of PTSD and PGD to clarify overlapping and distinct mechanisms and to strengthen the clinical utility of prediction models. Secondly, the substantial attrition from the initial cohort (n=287) to the final analytic sample (n=88) raises the possibility of selection bias. Participants who completed all assessments, including baseline blood tests, psychosocial questionnaires, and the 4-year follow-up PCL-5, may differ systematically from those lost to follow-up in terms of symptom severity, healthcare engagement, or socioeconomic factors. If individuals with more severe or unstable symptoms were less likely to complete longitudinal assessments, our model may not generalize well to harder-to-reach populations. Furthermore, the cohort comprises bereaved families from a single large-scale disaster in South Korea, which may limit generalizability to other trauma-exposed populations with different cultural, situational, or demographic characteristics. Third, probable persistent PTSD was operationalized using symptom status at the 4-year follow-up, which does not capture the full trajectory of chronicity. PTSD symptoms can fluctuate over time with periods of worsening and improvement; thus, the present model estimates risk of persistence at a single time point rather than characterizing fixed trajectory classes. Future studies with repeated intermediate assessments would help delineate trajectories more definitively. Additionally, psychosocial variables were derived from self-administered questionnaires rather than clinician-administered diagnostic interviews, which may introduce reporting biases. Class imbalance between the probable persistent (n=67) and probable remitted (n=21) groups, although addressed with SMOTE and stratified cross-validation, may still affect model robustness. Incorporating structured clinical interviews and external validation cohorts in future research would help corroborate and refine these findings. Fourth, while the core psychosocial predictors (GAD, PHQ, and ERR-1) showed high selection stability across bootstrap iterations, secondary features including HDL and coping subscales were less stable. These features were retained to provide a multifaceted bio-psycho-social perspective; however, their relative rankings should be interpreted with caution, and validation in larger, independent cohorts is warranted to confirm their incremental clinical utility. Finally, the absence of external validation with independent datasets constrains inferences about transportability and calibration across settings. Further study should assess PTSD and PGD together using longitudinal, multimodal measurement that combines self-report, clinician interviews and expanded biomarker panels. These predictions should be embedded in stepped-care trials to target modifiable risks such as insomnia, rumination, anxiety, and depression and to demonstrate real-world clinical utility.
In conclusion, this study highlights the importance of integrating psychological, biological, and social factors into predictive models for PTSD outcomes. The critical role of depression, anxiety, insomnia, rumination, social support and positive resources in distinguishing between prolonged and probable remitted outcomes identifies these psychosocial features as key predictors for PTSD persistence. Among the AI algorithms, LDA proved most effective, with high AUC, precision, and specificity, demonstrating its suitability for psychological and clinical datasets. The application of AI algorithms, particularly LDA, demonstrates the potential for advancing personalized mental health care, paving the way for more effective identification and management of individuals at risk for probable persistent PTSD.
Supplementary Materials
The Supplement is available with this article at https://doi.org/10.30773/pi.2026.0026.
Model performance metrics with bootstrap-based optimism correction
Standardized discriminant function coefficients for the final LDA model. Bar plot showing the standardized discriminant function coefficients for the seven features selected in the final LDA model. Positive coefficients indicate features associated with increased risk of probable persistent PTSD, while negative coefficients indicate features associated with probable remission. LDA, linear discriminant analysis; GAD, Generalized Anxiety Disorder-7; PHQ, Patient Health Questionnaire-9; ERR-1, Event-Related Rumination Inventory negative intrusive rumitation; ISI, Insomnia Severity Index; FSSQ, functional social support questionnaire; POREST, Positive Resources Test; HDL, high-density lipoprotein.
Feature selection stability across bootstrap iterations. Bar plot illustrating the selection frequency (%) of each candidate predictor across 1,000 bootstrap iterations. Features were ranked by their selection frequency to assess the robustness of the feature selection process. Core psychosocial predictors (GAD, PHQ, ERR-1) demonstrated high stability (>99%), while secondary features showed moderate to lower selection frequencies, reflecting potential overlap in predictive information. GAD, Generalized Anxiety Disorder-7; PHQ, Patient Health Questionnaire-9; ERR, Event-Related Rumination Inventory; FSSQ, functional social support questionnaire; ISI, Insomnia Severity Index; POREST, Positive Resources Test; B-COPE, Brief COPE subscales; HDL, high-density lipoprotein.
Notes
Availability of Data and Material
The datasets generated or analyzed during the study are available from the corresponding author on reasonable request.
Conflicts of Interest
The authors have no potential conflicts of interest to disclose.
Author Contributions
Conceptualization: Daun Shin, Byung-Joo Ham, So Hee Lee. Data curation: So Hee Lee. Formal analysis: Daun Shin, Beomgi So, Byung-Joo Ham. Funding acquisition: Jeong-Ho Chae, So Hee Lee. Investigation: Jeong-Ho Chae. Methodology: Daun Shin. Project administration: Jeong-Ho Chae, So Hee Lee. Resources: Byung-Joo Ham. Software: Beomgi So. Supervision: Byung-Joo Ham, So Hee Lee. Visualization: Daun Shin, Beomgi So. Writing—original draft: Daun Shin, Beomgi So. Writing—review & editing: Daun Shin, Byung-Joo Ham, So Hee Lee.
Funding Statement
This research was supported by Culture, Sports and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture, Sports and Tourism in 2025 (Project Name: Development of Technology for Mind Healing Service Platform Based on Humanities Program, Project Number: RS-2025-02308715, Contribution Rate: 100%)
Acknowledgments
We would like to express our sincere gratitude to all the participants who took part in this study.