Automatic Detection of Periodic Limb Movements in Sleep (PLMS), And Analysis of Their Relationship to Heart Rate Signals

Abstract

Periodic Limb Movement in Sleep (PLMS) are a sleep-related disorder of the limbs that increasingly more research has begun to associate with severe Cardiovascular Diseases (CVD). With that said, Polysomnography (PSG), followed by manual scoring, is the conventional approach being used to monitor the disorder. However, patient inconvenience, and the high costs associated with PSG, has probed the need for alternative screening tools to be developed. Moreover, due to the cumbersome and time-consuming nature of manually scoring for PLMS, more studies have begun to look into automated means of detecting PLMS. Hence, while one of the goals of the current thesis was to use the latest clinical specifications to develop an automated Periodic Limb Movement (PLM) detector, the other goal was to look into alternative signals to monitor PLMS. With that said, in the current thesis, an automated PLM detector was developed and tested on two datasets. In fact, the results were promising in that, correlation coefficients of 0.78 and 0.8, and absolute differences not greater than 9 and 6 (not including the extreme outliers) respectively, were found when comparing the clinical PLM scores with that of the automated algorithm’s PLM scores. Moreover, not only did the automated PLM detector compute PLM scores, it also provided us with PLM segmentation information, i.e., localization of PLM with respect to time. On the other hand, with regards to finding alternative signals to monitor PLMS, the etiology of PLMS was used in order to validate the use of relatively easily acquirable signals, such as Heart Rate (HR) signals, to monitor the condition. Moreover, core features were extracted from the HR signals and the PLM segmentation information from the developed PLM detector was used in order to perform individuaized classification between PLM and non-PLM segments (per subject). Although the results were promising in that, the percent of correctly identifying a given segment as PLM or non-PLM, using the HR features, across most of the subjects, i.e., especially those with PLM Index ≥ 15, were around and well above the 70% range, due to the possibility of other factors interfering with HR during sleep, a more immediate application of the observed PLMS vs HR distinction was, to be able to monitor the autonomic health of an individual, given their PLM information. Specifically, the latter was anticipated to be useful for studies looking into the relationship between PLMS and HR, and thus CVD, or more significantly, those looking into preventing CVD by treating PLM.