Design of pervasive systems for chronic sleep/brain disorders

Obstructive sleep apnea (OSA) is a highly prevalent sleep-related breathing disorder that is accompanied by major neurocognitive and cardiovascular comorbidities. The global prevalence of OSA is estimated to be nearly 1 billion people. It has been described as a "hidden health crisis" because around 80% of individuals with OSA syndrome go undiagnosed because they do not recognize the severity of the condition and therefore never reach out to a healthcare practitioner. In recent years, there has been a significant expansion in the development and use of consumer devices and technologies with advanced sensors that measure bio-signals accurately. OSA events trigger a sequence of responses that in particular affects the balance of the autonomic nervous system (ANS). It has been demonstrated that heart rate variability (HRV), which is a commonplace measurement in wrist-worn consumer sleep technologies (CSTs), constitutes a non-invasive and easy to derive signal that captures these autonomic manifestations reliably. Leveraging CSTs to screen for OSA opens the possibility to reach unidentified cases at a preclinic stage by providing an accessible and educational screening tool. Despite important advancements in this field in recent years, current OSA detection systems are limited because they are either feature-based or based on proprietary algorithms. Furthermore, these systems have only been tested on relatively small datasets. Deep learning (DL) methods are data-driven systems that have the potential to capture the complexity of sleep-disordered breathing (SDB) events, hereunder OSA, better than feature-based approaches. Validating such systems on raw data from CSTs will establish the reliability of CSTs as OSA screening tools.
This thesis, partitioned into three parts, presents data-driven-, automated classification systems that use artificial intelligence to explore the clinical potential of using CST as reliable sleep monitoring systems and OSA screening tools.
The first part presents a recurrent, electrocardiogram (ECG)-based SDB event detection system trained and validated on two large, cross-sectional clinical cohorts. Here, it was established that the model performed well to estimate the apnea-hypopnea index (AHI), i.e. the number of apneas per hour of sleep, with 85% accuracy, and with a correlation coefficient of R2 = 0.83. The model performance was robust across cohorts, across periodic leg movement index (PLMI) severity groups, and for subjects with previously reported incidents of cardiovascular disorders.
Part two presents a flexible DL model for sleep stage classification to explore the potential of using wrist-worn CSTs as reliable sleep monitoring tools. The model was designed with a strong temporal core that proved to capture long term-dependencies, as performance increased with increasing input segment size. Spectral preprocessing proved to make the model generalize well to an external test dataset, outperforming both low-resolution, raw, surrogate, and feature-based approaches.
The final part presents an improved version of the DL model from the second part, adapting it to predict both sleep stages and detect SDB events. SDB severity groups were estimated with a R2 = 0.54, R2 = 0.79, and R2 = 0.72 and Cohen’s k = 0.55, k = 0.66, and k = 0.57 for the pretraining-, internal-, and CST-based external test datasets, respectively. Pretraining the model on data from a large, heterogenous dataset increased the performance of the DL system substantially. The model generalized well to an external test dataset, despite the performance drop that was observed when the model was applied to datasets with photoplethysmography (PPG) signals recorded with CSTs.
In conclusion, the system performs well to classify both sleep stages and detecting SDB events when applied to recordings with high signal quality and to apneas associated with more pronounced hypoxia and autonomic responses. The system could work as a screening tool at a pre-clinic stage, targeting severe OSA cases, or as an out-of-clinic sleep monitoring tool for patients with OSA syndrome to track the progression of the condition and to enhance patient-clinician interaction.