|M.Sc Student||Kotzen Kevin Jacob|
|Subject||Sleep Architecture and Fragmentation Estimation from|
Photoplethysmography using Feature Engineering
and Deep Learning
|Department||Department of Biomedical Engineering||Supervisors||DR. Joachim Behar|
|ASSOCIATE PROF. Amir Landesberg|
|Full Thesis text|
Introduction: Sleep disorders are highly prevalent, affecting up to one-sixth of the global adult population. Sleep staging is an essential component of the diagnosis of sleep disorders. Sleep staging is traditionally measured using a specialized sleep study called polysomnography (PSG). PSG is expensive as it is usually performed in a clinical setting and requires a labor-intensive labeling process. Sleep stages are coupled to autonomous nervous system (ANS) activity, which in turn affect interbeat interval (IBI) variation of the heart. It has been shown that using IBI variation as a proxy for ANS activity one can estimate sleep stages. Most modern smartwatches can measure photoplethysmography (PPG) from which IBI can be derived. We hypothesize that it is possible to perform robust 4-class automated sleep staging from the raw PPG time series using modern advances in deep learning (DL).
Methods: We compared three machine learning (ML) approaches to sleep staging from PPG. The first approach used handcrafted features engineered (FE) from the PPG and a neural network classifier, the second approach used a derived time series (DTS) extracted from IBIs of the PPG as input to a DL classifier, and the third approach used the raw PPG as input to an advanced DL classifier. Models for the approaches are named Benchmark-FE (BM-FE), Benchmark-DTS (BM-DTS), and our new algorithm SleepPPG-Net. SleepPPG-Net consists of a residual convolutional network for automatic feature extraction and a temporal convolutional network to capture long-range contextual information. Using three publicly available databases, we train and benchmark the performance of SleepPPG-Net against BM-FE and BM-DTS models based on the best-reported state-of-the-art (SOTA) algorithms. The Sleep Heart Health Study Visit 1 (SHHS), totaling 5,758 unique patients, was used for model pretraining. The Multi-Ethnic Study of Atherosclerosis (MESA), consisting of 2,054 patients totaling 19,998 hours, was used for training and testing. The Cleveland Family Study (CFS) consisting of 320 patients totaling 3,057 hours, was used for evaluating generalizability on an external test database. Results: When benchmarked on a held-out test set of 204 MESA patients, SleepPPG-Net obtained a Cohen's Kappa (K) score of 0.75 against 0.69 for the best SOTA approach. SleepPPG-Net showed good generalization performance, obtaining a median K score of 0.74 after being fine-tuned to CFS using transfer learning. Contributions: This research makes the following contributions to the field; (1) it provides the first rigorous comparison of SOTA algorithms for sleep staging, (2) it introduces a new DL architecture, SleepPPG-Net, for sleep staging from the raw PPG, (3) it compares the sleep staging performance obtainable from electrocardiogram and PPG, (4) it evaluates the generalization to external test sets, (5) it provides a recommendation on what approach works best, and (6) it presents a novel PPG peak detector benchmarking tool and a new PPG morphological features toolbox. Perspective: Overall, SleepPPG-Net provides new SOTA performance. In addition, performance is high enough to open the path to the development of wearables that stage sleep accurately enough for usage in clinical applications such as the diagnosis of obstructive sleep apnea.