Interpretable machine learning for hikikomori screening: The adaptive HRI-15.

Abstract

Hikikomori, or prolonged social withdrawal, is an issue of global relevance. The HRI-15 is a brief tool for its assessment. This study enhances its utility by adding clinical thresholds to identify risk and person-centered clinical profiles, and by developing a machine-learning(ML)-based computerized adaptive test (CAT) to enable rapid screening. Data from a national survey of Italian adolescents (N = 8,755) were used to conduct ROC analysis and latent profile analysis (LPA). The findings indicated that a score of ≥ 42 achieved optimal classification performance, and four profiles with distinct meanings and systematic differences in anxiety, depression, impulsivity, and risk behaviors were identified. A multivariate conditional inference tree was estimated to develop an ML-based adaptive version of the instrument. The CAT reduced administered items by 53% (7.03/15), accurately reproducing full-length scores (r = .77-.997) and the corresponding classification; alignment was assessed with latent transition analysis (entropy = .96). The procedure was integrated into an application supporting administration, scoring, and reporting (score, risk, profiles, graphs). Combining a cross-validated cutoff and clinically useful profiles with CAT and its automated administration and reporting application strengthens triage and personalization, enabling large-scale, multidomain screening and earlier intervention for hikikomori risk.