Pilot study: can machine learning analyses of movement discriminate between leg movements in sleep (LMS) with vs. without cortical arousals?

Purpose Clinical and animal studies indicate frequent small micro-arousals (McA) fragment sleep leading to health complications. McA in humans is defined by changes in EEG and EMG during sleep. Complex EEG recordings during the night are usually required to detect McA-limiting large-scale, prospective studies on McA and their impact on health. Even with the use of EEG, reliably measuring McA can be difficult because of low inter-scorer reliability. Surrogate measures in place of EEG could provide easier and possibly more reliable measures of McA. These have usually involved measuring heart rate and arm movements. They have not provided a reliable measurement of McA in part because they cannot adequately detect short wake periods and periods of wake after sleep onset. Leg movements in sleep (LMS) offer an attractive alternative. LMS and cortical arousal, including McA, commonly occur together. Not all McA occur with LMS, but the most clinically significant ones may be those with LMS [ 1 ]. Conversely, most LMS do not occur with McA, but LMS vary considerably in their characteristics. Evaluating LMS characteristics may serve to identify the LMS associated with McA. The use of standard machine learning approaches seems appropriate for this particular task. This proof-of-concept pilot project aims to determine the feasibility of detecting McA from machine learning methods analyzing movement characteristics of the LMS. Methods This study uses a small but diverse group of subjects to provide a large variety of LMS and McA adequate for supervised machine learning. LMS measurements were obtained from a new advanced technology in the RestEaZe™ leg band that integrates gyroscope, accelerometer, and capacitance measurements. Eleven RestEaZe™ LMS features were selected for logistic regression analyses. Results With the optimum logit probability threshold selected, the system accurately detected 76% of the McA matching the accuracy of trained visual inter-scorer reliability (71–76%). The classifier provided a sensitivity of 76% and a specificity of 86% for the identification of the LMS with McA. The classifier identified regions in sleep with high versus low rates of LMS with McA, indicating possible areas of fragmented versus undisturbed restful sleep. Conclusion These pilot data are encouraging as a preliminary proof-of-concept for using advanced machine learning analyses of LMS to identify sleep fragmented by McA.