Circadian rhythm sleep-wake disorders (CRSWDs) arise from misalignment between an individual's sleep-wake timing and endogenous circadian physiology, with substantial consequences for sleep health and overall well-being. In the era of sleep-omics and digital phenotyping, machine learning (ML) offers powerful tools for modelling complex, high-dimensional circadian data; however, its application to CRSWD prediction remains under developed. This systematic review aims to synthesize current evidence on the use of ML approaches for circadian phase estimation and CRSWD classification in humans. We have conducted a systematic search of PubMed, ScienceDirect, PsycINFO, Cochrane Library, and Google Scholar for studies published between January 2011 and May 2025. Eligible studies focused on sighted adults where ML algorithms predicted circadian phase or classified CRSWDs. Data sources included sleep diaries, actigraphy, genomics, and biological markers. To maintain model homogeneity and account for physiological differences, studies involving pediatric populations, neurodegenerative comorbidities, or blind individuals were excluded. A total of 22 studies met the inclusion criteria. Across studies, ML models demonstrated promise for circadian phase estimation and CRSWD classification using diverse data modalities, including wearable-derived time-series data and multimodal features. However, the evidence base remains limited by small sample sizes, methodological heterogeneity, limited external validation, and inconsistent reporting standards. Overall, current findings suggest that while ML-based approaches hold significant potential for advancing CRSWD detection and personalized circadian medicine, more robust, standardized, and generalizable models are required. Future research should prioritize larger multimodal datasets, transparent validation strategies, and clinically translatable tools to support scalable CRSWD screening and diagnosis.
