Machine learning clustering of insomnia subtypes using digital phenotype data

초록

Introduction: Insomnia diagnosis relies heavily on self-reports, leading toambiguity in classification. Digital phenotyping, which analyses beha-vioural patterns from digital devices, offers a novel approach to insomniasubtyping. This study aims to explore data-driven insomnia subtypingusing digital phenotypes and their relationship with clinical characteristics.Method: Seventy-three adults (aged 19–70) with Insomnia SeverityIndex scores ≥15 were recruited from Korea University Anam Hospi-tal. Digital phenotypes including sleep quality, life patterns, and envi-ronmental factors were recorded via a mobile application for 4 weeks.K-means clustering was performed using 16 variables (e.g., steps,exercise time, movement distance). Between-group analyses of demo-graphic and clinical characteristics were conducted using ANOVA, t-tests, or chi-square tests as appropriate.Results: K-means clustering identified three subgroups (n = 32,31, 10). Steps, exercise time, and movement distance were the mostdistinctive features. Cross-validation achieved 80% accuracy. SHAPvalue analysis confirmed the importance of movement-related fea-tures. Among 28 clinical characteristics, only the Dysfunctional Beliefsand Attitudes about Sleep (DBAS) score differed significantly betweenclusters (p = 0.04). Post-hoc t-tests revealed significant differencesbetween clusters (C0 = C1, C1 = C2, C0 < C2; p = 0.01).Conclusion: Despite limited parameters and sample size, digital phe-notyping effectively clustered insomnia subtypes, with movement-related features emerging as crucial. The significant difference inDBAS scores between clusters suggests a potential link between digi-tal phenotypes and sleep-related cognitions. These findings indicatethat digital phenotyping could provide valuable insights for persona-lised insomnia treatment approaches.