Machine learning to reveal hidden risk combinations for the trajectory of posttraumatic stress disorder symptoms

PTSD symptom trajectories after disasters are heterogeneous and only weakly explained by single risk factors. Prior longitudinal studies showed modest effects of demographics and exposure on symptom trends and highlighted interactions among risk and protective factors. However, traditional approaches typically preselect variables based on bivariate associations, potentially missing factors that exert influence only through interactions (“hidden risk components”). Given the combinatorial explosion and multiple-testing burden, comprehensive detection of risk combinations has been infeasible. MP-LAMP, an algorithm that prunes untestable and dependent combinations while rigorously controlling familywise error, enables exhaustive, interaction-aware discovery. This study focuses on long-term prognosis of PTSD symptoms following the Great East Japan Earthquake, aiming to identify combinational risk factors, including hidden components, that predict 8-year PTSD trajectories beyond baseline severity.

Previous disaster studies reported small effect sizes of single predictors for PTSD prognosis; for example, middle age and low income explained about 2.1% of variance in trajectories, and exposure predicted onset but not trends. Interaction effects have been observed: social support impacted females more; excessive alcohol use affected males’ symptom exacerbation; family loss influenced younger more than older individuals. Work considering combinations suggested hidden long-term impacts of war exposure. Yet most prior work examined pairwise interactions among a limited set of factors due to computational constraints, potentially overlooking higher-order interactions and hidden components. The present work addresses this gap by using MP-LAMP to comprehensively test combinations without preselection by bivariate significance.

Design and population: Community-based longitudinal cohort from the Shichigahama Health Promotion Project targeting residents whose homes were totally collapsed or severely damaged by the March 11, 2011 Great East Japan Earthquake/Tsunami. Surveys were administered annually; this analysis used data from 2011 (year 1), 2012 (year 2), and 2018 (year 8). Of 2,478 eligible adults (≥18 years), 1,791 participated in year 1; 1,173 in year 2; and 636 in years 1, 2, and 8. After excluding respondents with >20% missing on IES-R or risk factors, and >50% missing overall, the final analytic sample was 624. Measures: Outcome was the Impact of Event Scale-Revised (IES-R; 0–88) at year 8 adjusted for baseline (year 1) IES-R, termed the PTSD trajectory score (residualized change reflecting variance not explained by baseline). Predictors were 61 potential risk factors collected mainly in year 1: sociodemographics (age, sex, employment), lifestyle (smoking, alcohol, daily walking/sitting/sleeping time), clinical history, K6, AIS, LSNS-6; and disaster-related experiences and economic changes (from year 2: evacuation, witnessing tsunami, life threat, witnessing threats to others, bereavement, decreased income/work volume). Data handling and preprocessing: Missingness among IES-R items and predictors was 0.5% and 2.9%, respectively. After checking for missing-data bias, missing values were imputed nonparametrically using the missForest R package to obtain a complete dataset required by LAMP. Because MP-LAMP requires binary predictors, variables were binarized: established cutoffs were used for K6 (5/6; 12/13), AIS (5/6), and LSNS-6 (11/12). Other ordinal (≥3 levels) and continuous variables were discretized into three approximately equal-frequency bins (infotheo R package), then dichotomized by designating the highest or lowest tertile as the risk group and the remaining as nonrisk to maintain computational feasibility. Analytic approach: MP-LAMP (https://github.com/tsudalab/mp-lamp) was used for comprehensive detection of statistically significant combinations. LAMP prunes combinations that cannot achieve significance due to low support and avoids redundant tests for completely dependent combinations, employing frequent itemset mining and a calibrated Bonferroni correction to strictly control the familywise error rate. Main analyses used unadjusted predictors and the PTSD trajectory score; an additional analysis used the trajectory score adjusted for age and sex to assess robustness. For comparison, bivariate linear regressions (age- and sex-adjusted) of each single predictor on the trajectory score were conducted with Bonferroni correction. Associations between identified combinations and the trajectory score were further evaluated via Mann–Whitney U tests. Statistical significance was defined as P<0.05.

- Baseline contribution: Year-1 IES-R explained 23.5% of variance in year-8 IES-R; remaining variance was examined via the PTSD trajectory score. - Bivariate analyses: After multiple-testing correction, no single demographic or trauma-exposure factor showed a significant association with the PTSD trajectory score; older age, female sex, and higher trauma exposure were associated with higher baseline IES-R but not with trajectory. - MP-LAMP combinational findings: 56 significant risk combinations were detected, involving 15 unique component variables at least once. Each significant combination contained at least one component without a significant bivariate association (raw P>0.05), indicating hidden risk components. Interactions among components within significant combinations were enriched: mean (SD) counts of interactions with P<0.05 and P<0.01 were 4.9 (2.9) and 2.5 (1.3), exceeding 95% CIs from 100,000 random combinations (1.2–1.9 and 0.3–0.7, respectively). - Strongest combination: Unemployment + walking <30 min/day + sitting/napping <3 h/day + evacuation without preparation showed the strongest association with poorer PTSD trajectory (adjusted P=2.2e-4; raw P=3.1e-9) and a much larger effect on year-8 IES-R than any single component. Short resting time alone was not significant (raw P=0.055) but interacted significantly with short walking time (P=1.2e-3), and the interaction effect increased further when combined with unemployment and unprepared evacuation (P=9.7e-5). - Robustness to adjustment: The age- and sex-adjusted analysis identified combinations largely overlapping with the main results; the top 10 combinations from the main analysis remained significant, and all 15 combinations in the adjusted analysis were also significant in the main analysis. - Variance explained: The largest variance explained in the trajectory score by a combination (unemployment, short walking time, short resting time, evacuation without preparation, life-threatening experience, decreased income) was 8.5%. The strongest single predictors (poor physical condition and decreased work) each explained about 2.0%. - Gender interactions: Although female sex did not affect the trajectory in bivariate analysis (P=0.45), it interacted with decreased income (P=2.7e-3), poor physical condition (P=8.1e-3), and older age (P=0.025), appearing in some significant combinations.

By leveraging MP-LAMP to examine all feasible combinations while controlling the familywise error rate, the study uncovered significant multi-factor risk patterns for long-term PTSD symptom trajectories that were invisible to single-factor analyses. Findings directly address the hypothesis that interactions among sociodemographic, lifestyle, and trauma-related factors underpin prognosis: every significant combination included at least one hidden component lacking a bivariate association, and interaction enrichment analyses supported genuine synergistic effects, including higher-order interactions beyond pairs. Clinically, results suggest that traumatic exposure alone may not elevate long-term risk unless coupled with social/employment and lifestyle adversities (e.g., unemployment and low activity with unprepared evacuation). This reframes risk assessment toward integrated profiles capturing trauma, social context, and behavior. The observed gender interactions further indicate that economic and health stressors may differentially impact women’s recovery trajectories. Overall, combinational risk modeling improves explanatory power (up to 8.5% variance by combinations vs ~2% by single factors) and can inform targeted surveillance and interventions for high-risk subgroups years after disasters.

A comprehensive combinational approach using MP-LAMP substantially increased power to detect predictors of long-term PTSD trajectories and revealed hidden risk components present only through interactions. The strongest combinations integrated employment status, physical activity/rest, and disaster-response factors, outperforming any single predictor. These results advocate for interaction-aware risk profiling in psychiatric prognosis and support future research applying comprehensive combination detection across diverse traumas and populations to validate and extend these findings.

- Sample size was modest (n=624), a common constraint in post-disaster cohorts, which may limit precision and generalizability. - Current MP-LAMP implementation does not natively adjust for covariates; authors mitigated this by analyzing a trajectory score adjusted for age and sex, with largely consistent results, but residual confounding cannot be fully excluded. - External validation is lacking; significant combinations require replication in independent cohorts, different ethnicities, and varied trauma types. - Limited measurement occasions for the long-term analysis (primarily years 1 and 8) and lack of detailed information on intervening exposures may complicate interpretation of long-term trajectories.