Posttraumatic stress disorder (PTSD) is a complex condition with varied recovery trajectories. While individual risk factors for PTSD have been studied, their effects are often weak and insufficient to explain the observed diversity in symptom trajectories. Previous research suggests that interactions among multiple risk and protective factors significantly influence PTSD prognosis, but most prior studies had limitations in their approach to identifying these interactions. Comprehensive combination detection studies examining all potential interactions among risk factors are potentially more powerful, but have been hindered by high computational costs and stringent multiple-testing corrections. This study addresses these challenges by employing a novel machine learning approach to uncover hidden risk combinations related to PTSD symptom trajectory. The Massive Parallel Limitless-Arity Multiple-testing Procedure (MP-LAMP), developed to detect significant combinations of risk factors, was utilized in a longitudinal study of individuals severely affected by the Great East Japan Earthquake, to analyze a comprehensive set of potential risk factors and their interactions on PTSD trajectories eight years post-disaster. Understanding these combinations has significant clinical implications, potentially leading to more effective strategies for identifying high-risk individuals and tailoring interventions to improve long-term outcomes.

Existing literature on PTSD prognosis after disasters reveals a complex interplay of factors, but typically focuses on individual risk factors or limited pairwise interactions. Studies using bivariate analysis have shown weak associations between demographic characteristics (age, income, gender), traumatic experiences, and PTSD symptom trajectories. For instance, Kessler et al. (2008) found middle age and low income to be slightly associated with PTSD trends post-hurricane, explaining only 2.1% of the variance. Similarly, Adams and Boscarino (2006) found exposure to stressful events to be predictive of PTSD onset, but not its trajectory. However, studies such as Andrews et al. (2003), Soo et al. (2011), and Bokszczanin (2007) highlight the importance of interactions. For instance, social support's positive impact on PTSD prognosis is stronger for females, while excessive alcohol intake exacerbates symptoms in males. Furthermore, Drožđek et al. (2020) demonstrated the hidden long-term impacts of combined war-related exposures. Although these studies point towards the value of considering interactions, a limitation is the pre-selection of candidate risk factors based on prior associations. This selection bias could exclude factors that only reveal their importance when combined with others, leading to an incomplete understanding of the complex interplay of risk factors influencing PTSD prognosis.

This study utilized data from the Shichigahama Health Promotion Project, a longitudinal health survey of residents whose houses were destroyed or severely damaged by the Great East Japan Earthquake. The analysis included 624 subjects who completed surveys in 2011, 2012, and 2018. The Impact of Event Scale-Revised (IES-R) was used to measure PTSD symptoms, with the 8th-year IES-R score adjusted for the 1st-year score (PTSD trajectory score) serving as the outcome variable. Sixty-one potential risk factors were included, covering sociodemographics, lifestyle, traumatic experiences, and clinical information. Missing data were imputed using the missForest package in R. The core methodology involved the application of MP-LAMP, a parallel computing algorithm that efficiently identifies statistically significant combinations of risk factors. Unlike traditional methods that assess individual risk factors or limited combinations, MP-LAMP evaluates all possible combinations, effectively accounting for interactions. MP-LAMP operates by reducing computational costs and mitigating the multiple testing problem by ignoring combinations that cannot be statistically significant and removing redundant combinations. For the MP-LAMP analysis, continuous and ordinal variables were converted into binary variables using appropriate cutoffs or discretization methods. The results of MP-LAMP were compared with those of conventional bivariate analyses (linear regression adjusted for age and sex) to highlight the differences between single-factor and combinatorial analyses. The Mann-Whitney U test was used to evaluate associations between potential risk combinations and PTSD trajectory scores. All analyses were conducted in R, and p<0.05 was considered statistically significant.

While conventional bivariate analyses failed to identify significant predictors of PTSD trajectory scores, the MP-LAMP analysis revealed 56 significant combinations of risk factors, involving 15 independent variables. The strongest association (adjusted p = 2.2 × 10⁻¹⁰, raw p = 3.1 × 10⁻⁹) was observed for the combination of unemployment, walking less than 30 minutes/day, short resting time (less than 3 hours/day), and evacuation without preparation. Importantly, many of the identified risk factors (10 out of 15) showed no significant association with the outcome variable in the bivariate analyses. These 'hidden risk components' only revealed their significance through their interactions with other factors within the combinations. The average number of significant interactions (p < 0.05 and p < 0.01) within the identified combinations was significantly higher than expected by chance, further underscoring the importance of the identified interactions. An additional analysis adjusted for age and sex confirmed the robustness of the key findings, showing substantial overlap between the significant combinations identified in the main and adjusted analyses. The combination of factors identified explained 8.5% of the variance in the PTSD trajectory score, which is a significant increase compared to previous studies focusing on individual factors, thereby demonstrating the added predictive value of the combinatorial approach.

The findings demonstrate that a comprehensive, combinational approach to risk factor analysis offers substantially greater power in identifying significant predictors of PTSD symptom trajectories than traditional bivariate analyses. The identification of 'hidden risk components' highlights the limitations of focusing solely on individually significant risk factors. The interaction among factors, particularly involving those that are not significant on their own, significantly contributes to the risk profile. The combination of unemployment, limited physical activity, insufficient rest, and unprepared evacuation represents a clinically relevant profile indicative of increased vulnerability to prolonged PTSD symptoms. The interaction effects point towards the importance of lifestyle factors, work-related stress, and preparation for emergencies in mediating the long-term recovery from traumatic events. These findings necessitate a shift towards considering the cumulative effect of risk factors and their interactions, rather than solely focusing on individual risk factors when assessing risk for and developing interventions for PTSD.

This study successfully employed MP-LAMP to identify significant combinations of risk factors for long-term PTSD symptom trajectory, revealing substantial interactions among previously overlooked factors. The identification of 'hidden risk components' strongly suggests that comprehensive combinational approaches are crucial for accurately predicting and managing PTSD. Future research should focus on validating these findings in independent cohorts and exploring the specific mechanisms driving these identified interactions. Furthermore, investigation into the generalizability of these risk combinations across various trauma types and populations is needed.

The primary limitation of this study is the relatively small sample size (624 participants). While statistically significant results were obtained with MP-LAMP, future studies with larger samples are required to further validate the findings and improve generalizability. Another limitation is the reliance on self-reported data, which might be subject to recall bias. Furthermore, while the MP-LAMP analysis was not explicitly adjusted for confounders, an additional analysis adjusted for age and sex produced consistent results, suggesting minimal influence from confounders. Lastly, replication studies in independent samples with diverse demographics and trauma exposure are crucial for verifying the external validity and generalizability of the results.