Musashino University

Hikikomori, characterized by prolonged social withdrawal and isolation, is often comorbid with psychiatric conditions including depression and anxiety. While cross-sectional studies consistently link hikikomori with depressive symptoms, the directionality of this relationship remains unclear because longitudinal evidence is lacking.

This study examined the longitudinal association between depressive symptoms and hikikomori behavioral tendencies, focusing on both the topography (observable features) and functions of hikikomori behaviors, as conceptualized in operant conditioning theory.

A total of 497 individuals aged 20-64 years who met established criteria for hikikomori were surveyed three times over 12 months. Hikikomori behaviors and perceived functions were assessed using the Adaptive Behaviors Scale for Hikikomori and the Hikikomori Functional Assessment Scale. A random intercept cross-lagged panel model was employed to disentangle the within- and between-person effects, controlling for sex, self-compassion, and social support.

Depressive symptoms predicted increased hikikomori behavior over time, whereas hikikomori behavior did not predict subsequent increased depressive symptoms. A bidirectional association was found between depressive symptoms and hikikomori maintained by social negative reinforcement (e.g., avoidance of social interaction). In contrast, hikikomori behaviors maintained by intrapersonal positive reinforcement were not significantly associated with depressive symptoms.

These findings suggest that depressive symptoms were prospectively associated with the development and maintenance of hikikomori, particularly when driven by avoidance of aversive social experiences. Interventions targeting depressive symptoms and maladaptive reinforcement processes, such as behavioral activation, may help interrupt this cycle. Future studies should investigate other psychiatric comorbidities and test these findings in different cultural contexts.

Interpretable machine learning is increasingly used in oncology, yet feature attributions from supervised models (e.g., Random Forest, XGBoost) can be unstable and bias-prone when grounded solely in SHAP explanations. We contrast target-prediction accuracy with the reliability of feature-importance estimates and assess stability via feature-elimination tests on TCGA data (705 samples, 1936 features). While supervised models achieved modest gains (Random Forest ROC-AUC: 0.8851 to 0.8865; XGBoost: 0.8681 to 0.8695), their feature-selection stability was low (6/10 and 3/10 retained, respectively). Unsupervised and non-target methods were markedly more stable: Feature Agglomeration, Highly Variable Gene Selection, and Spearman retained 10/10 features with unchanged performance (0.8823, 0.8823, 0.8766). We recommend combining unsupervised criteria with causal design and external validation to mitigate model-specific biases.