Predicting daytime sleepiness from ECG-based respiratory rate using deep learning

Emmi Antikainen1, Rana Zia Ur Rehman2, Teemu Ahmaniemi1, Meenakshi Chatterjee3
1VTT Technical Research Centre of Finland Ltd., 2Newcastle University, 3Janssen Research & Development


Abstract

Introduction: Fatigue and sleep disturbances are a common problem in chronic diseases and can contribute to daytime sleepiness. Daytime sleepiness further impacts the activities of daily living, ultimately deteriorating the quality of life. Hence, sleepiness is also an important aspect to evaluate when developing new therapies. Typically, daytime sleepiness is measured by subjective questionnaires such as the Karolinska Sleepiness Scale (KSS), while there is a need for development of more objective measures. This study employs a wearable sensor with ECG and accelerometer to measure respiratory rate. Inspired by the intuitive association between yawning and daytime sleepiness, we explore whether sleepiness can be predicted from continuously monitored respiratory rate using a deep learning approach.

Methods: The analysis comprised a total of 82 participants from four different sites, including healthy volunteers (N=26) and patients with immune-mediated inflammatory diseases (IMID, N=42), and neurodegenerative diseases (NDD, N=14). Each participant wore the patch-like sensor for 3–12 days, while responding to KSS questionnaires. During this time, participants were asked to report their sleepiness with the KSS questionnaires three times a day. A 1-dimensional convolutional neural network (CNN) with dropout was trained for binary classification between non-sleepy (KSS levels from “extremely alert” to “rather alert”) and sleepy (KSS levels from “neither alert nor sleepy” to “extremely sleepy”) status, using respiratory rate time-series as input data. The classification result was validated using 10-fold cross-validation, so that each participant only appeared in a single fold (training and validation on different participants).

Results: The suggested CNN model achieved an average accuracy of 82.3% (55.4–96.8%), sensitivity of 84.6% (66.7–98.0%), and specificity of 79.8% (48.8–97.6%). Using wearable sensors combined with deep learning methods, our results suggest that respiration rate is a predictor of daytime sleepiness.