Characterization of insomnia at home using digital tools and machine learning methods from the combination of objective and subjective criteria

Abstract

Insomnia is one of the most common sleep disorders, degrading quality of life and raising the risk of accidents and of cardiovascular and metabolic disease. Its assessment is split between objective measures such as sleep latency, wake after sleep onset, and total sleep time, obtained from polysomnography (PSG) or actigraphy, and the subjective complaint captured by questionnaires such as the Insomnia Severity Index (ISI). The two correlate poorly: many who report disturbed sleep meet no objective criteria, and some who sleep normally by self-report do. Insomnia is diagnosed from the subjective complaint, actigraphy tracking sleep-wake rhythms, with PSG reserved for detecting sleep abnormalities and comorbidities; this reference examination is costly, depends on scarce expert scorers, and covers one or two laboratory nights, so it reflects neither the home environment nor the night-to-night variability that characterizes the disorder. Connected sleep-tracking devices used at home offer a complementary route. The under-mattress Sleep Analyzer and the wrist-worn ScanWatch record activity and cardiorespiratory markers continuously over long periods and, when paired with questionnaires, support an assessment that integrates objective and subjective criteria. This thesis investigates how such tools, combined with machine learning, can characterize insomnia in ecological conditions and search for homogeneous subgroups of the disorder. The work comprises four contributions, each an independent publication. The first is a systematic review of digital tools for insomnia detection and characterization, which maps the signals, datasets, models, and validation practices in the literature and identifies recurrent weaknesses, including small datasets, poorly documented reference labels, and subject-level data leakage that inflates reported performance. The second evaluates the Sleep Analyzer sleep-stage algorithm against PSG in an at-home study. The third designs and runs the SLEEPI protocol, an ecological longitudinal study pairing objective device recordings with repeated questionnaires, among them the ISI, PHQ-9, and GAD-7, with labels from DSM-5 criteria. The fourth is a prospective analysis of the longitudinal device data that searches for insomnia subgroups in the objective sleep and activity features. Under unsupervised clustering these features form a single continuum rather than discrete subtypes, and the objective signal that tracks self-reported severity is weak, so the analysis reframes the objective phenotype as graded and points to within-person night-to-night variability as a dimension that longitudinal home monitoring can measure and that warrants dedicated future investigation. The thesis thus argues that longitudinal, at-home, multimodal monitoring is a practical path toward a more individualized characterization of insomnia than current assessment methods allow.