RBDGuard

REM Sleep Behavior Disorder (RBD) is a significant predictor of neurodegenerative diseases such as Parkinson’s Disease, but its diagnosis currently requires resource-intensive polysomnography (PSG) studies, limiting widespread screening. This paper presents RBDGuard, a novel deep learning framework that enables RBD detection using only single-lead electrocardiogram (ECG) data by leveraging both unlabeled and labeled datasets through unsupervised pre-training with supervised fine-tuning. My approach capitalizes on the increasing accessibility of ECG monitoring through wearable devices to offer an easy and inexpensive early diagnostic tool for sleep disorders. RBDGuard's novel architecture employs bidirectional LSTM layers for unsupervised pre-training followed by dense layers with dropout for supervised classification. RBDGuard utilized 377,622 unlabeled 30-second ECG segments from 9,174 subjects and 39,316 labeled 30-second ECG segments from 37 subjects (22 RBD, 15 control) across three public databases. Compared to the exclusively supervised model, RBDGuard achieved significant improvements in accuracy (from 80% to 97%) and reduction in misclassification (a 75% decrease in false positives and 95% decrease in false negatives), reducing both missed diagnoses and unnecessary referrals. Through extensive testing, I demonstrated that the model consistently maintained accuracies of 97%-98% across different pre-training datasets, outperforming the supervised-only approach and indicating its adaptability to different clinical settings. These promising results establish RBDGuard as an accurate, efficient alternative to traditional PSG studies, paving the way for widespread early detection of RBD using unsupervised learning with ECGs. This advancement enables cost-effective, accessible early diagnosis of neurodegenerative diseases, transforming medical care and improving millions of lives.