Continuous measurement of sleep, sedentary behavior, and physical activity levels from accelerometer data using robust algorithms and practical sensing systems
Binod Thapa-Chhetry · 2024
The National Heart, Lung, and Blood Institute and the CDC stress the importance of sufficient sleep and physical activity for health. However, a significant portion of U.S. adults do not meet recommended guidelines. Researchers need more accurate, ongoing monitoring of sleep and activity levels for the population, and large health surveys like the National Health and Nutrition Examination Survey and UK Biobank now collect extensive accelerometer datasets that provide an opportunity. Traditional machine learning algorithms struggle to analyze this data because they depend on short-term data patterns and do not adequately account for sequential and long-term dependencies. To inform public health decisions and intervention strategies more effectively, researchers need models capable of analyzing these datasets more accurately, identifying daily activity patterns, and providing a full 24-hour behavioral overview.In this dissertation, I introduce two methods (i.e., SPAN and SPADS) for accelerometer sensor data-based sleep and PA level classification. SPAN (Sleep Physical Activity level, and Non-wear classification) uses wrist accelerometer sensor data and SPADS (Sleep and Physical Activity levels classification using Dual-Sensor data) fuses potentially complementary information in data from sensors placed on non-wrist sites (i.e., ankle, thigh, and waist) to data from the wrist to improve classification performance. Both models leveraged temporal convolutional network (TCN), bidirectional long short-term memory (BiLSTM), and transformers to learn local, sequential, and global patterns in temporal accelerometer data. Both models were trained on first-of-a-kind multi-day 24-hour data with ground-truth labels based on Polysomnography (for sleep) and front-facing camera-based annotations (for waking behaviors). Internal validation results showed that both models performed better than their respective baseline models. The validation of SPAN across four datasets underscored its architectural strengths. Compared to TCN-only models, SPAN's utilization of BiLSTM layers enhanced its ability to accurately distinguish behaviors with similar movement patterns, especially sleep and non-wear periods, which are often conflated. Integrating a transformer layer post-TCN highlighted the model's proficiency in capturing global contextual cues, although it also revealed the limitations of relying on global context without considering sequential information. The independent test dataset included five diverse and representative datasets. SPAN outperformed four currently available methods for sleep and PA level classification. SPAN's sleep predictions aligned well with self-reported sleep measures, displaying its versatility across different demographic groups and among individuals with sleep disorders. This evidence supports SPAN's capability as a more accurate tool for passive sleep behavior measurement compared to existing methods. Similarly, SPADS showcased a consistent advantage over traditional data fusion methods. By integrating data from both the wrist and ankle, SPADS achieved a modestly superior classification of behaviors across all categories, particularly for activities that involve subtle movements, such as distinguishing between various intensities of physical activities. The performance improvement in sleep and PA level classification based on the two proposed models could also be clinically meaningful, as the additional precision in capturing and categorizing behaviors could inform more personalized health interventions and contribute to a deeper understanding of the interplay between various activities and overall well-being. This dissertation's work underscores the critical necessity for behavior classification models to be trained on diverse datasets that encompass a wide range of behaviors, ensuring their robustness and generalizability. Such training enables the models to integrate local, sequential, and long-range contextual data effectively, which is pivotal for a comprehensive understanding of human activities. Moreover, the adaptability of algorithms is highlighted as essential, allowing models to retain accuracy when applied to new and unseen datasets, an attribute that lays the groundwork for future advancements. The research conducted with the SPAN and SPADS models indicates a shift from a singular focus to an integrated, multi-contextual approach. This shift is grounded in the imperative for multi-dimensional data integration, the tailoring of algorithms to specific demographic patterns, a commitment to non-invasive measurement techniques, and a rigorous validation process against clinical standards. These models' insights offer a strategic roadmap for refining behavior classification algorithms, stressing the balance of recent methodological advancements, ease of use, and economic feasibility as fundamental aspects in the ongoing development of wearable monitoring systems. --Author's abstract