On Monitoring Brain Health from the Depths of Sleep: Feature Engineering and Machine Learning Insights for Digital Biomarker Development

Abstract

Background: Single-channel sleep electroencephalography (EEG) is a promising technology for creating cost-effective and widely accessible digital biomarkers for monitoring brain health. Sleep, notable for its numerous connections to brain health, is of particular interest in this context. Indeed, several of the best studied and widely recognized risk factors for neurodegenerative disease are also connected to aspects of sleep physiology, including biological sex, hypertension, diabetes, obesity/metabolic dysregulation, and immune system dysfunction. In this study, we utilize the unique signal characteristics of slow wave sleep (SWS) oscillatory events as features in machine learning models to predict underlying biological processes that are highly relevant to brain health. Our objective is to establish a foundation for algorithms capable of effectively monitoring physiological processes in sleep that directly and indirectly inform brain health using single-channel sleep EEG as a functional metric of brain activity. Methods: Utilizing data from the Cleveland Family Study, we analyzed 726 overnight polysomnography recordings to extract features from slow waves and adjacent oscillatory events. Advanced signal processing and machine learning techniques, including random forest models, were employed to engineer features and predict health-related outcomes such as age, cerebrovascular risk factors, endocrine functions, immune system activity, and sleep apnea. Results: Our models demonstrated significant predictive capability for several outcomes, including age (R2 = 0.643, p < 0.001), sex classification (area under the receiver operator characteristic (AUROC) curve = 0.808), diabetes and hypertension diagnosis (AUROC = 0.832 and 0.755, respectively). Significant predictions were also modeled for metabolic/endocrine functions (including blood concentrations of IGF-1, leptin, ghrelin, adiponectin, and glucose), and immune markers (including IL-6, TNF-alpha, and CRP). In addition, this approach provided successful predictions in regression modeling of BMI and both regression and classification of sleep apnea. Discussion: This study demonstrates the potential of using features from oscillatory events in single-channel sleep EEG as digital biomarkers. These biomarkers can identify key health and demographic factors that both affect brain health and are indicative of core brain functions. By capturing the complex interactions of neural, metabolic, endocrine, and immune systems during sleep, our findings support the development of single-channel EEG as a practical tool for monitoring complex biological processes through metrics that originate in brain physiology. Future research should aim to refine these digital biomarkers for broader home-based applications that may utilize inexpensive "wearable" devices to provide a scalable and accessible tool for tracking brain health-related outcomes.