Abstract
Continuous monitoring of health parameters, especially in real-world out-of-hospital settings, is vital for patients with chronic diseases in preventing acute and hazardous health outcomes. Sensing and machine learning (ML) can facilitate continuous health tracking to reduce such risks. For example, real-time sensing-based prediction of the agitated behavior in dementia patients can prevent harmful escalations, similarly, inference-based monitoring of the respiratory function in asthmatics can prevent asthma attacks.
State-of-the-art sensor-ML researches for health monitoring struggle to transfer in-lab performance to the real-world deployments, as models built with in-lab or limited duration “snapshot” data often fail to generalize temporally and spatially to real-world beyond-training scenarios. But continuous long-term health monitoring data from real-world settings are rare for most health applications and are challenging to acquire due to hurdles in system reliability and data availability, along with user compliance and usability. Challenges in model design with such large data range from inter-disease, inter-patient, and intra-patient variations among the health variables to signal noise and missing data. These challenges are aggravated by the lack of reliable ground truth from real users for such long-term data as well as by the lack of generalizability of the model performance beyond training scenarios.
This dissertation addresses these challenges for wearable sensor-ML systems in the acquisition and utilization of real-world long-term data for human behavior and physiology learning. To achieve robust data collection in real-world settings, a novel wearable-edge platform, named behavioral and environmental sensing and intervention (BESI), is proposed and implemented to facilitate usability, unobtrusiveness, reliability and data availability. Using such real-world data, this work explores learning methods for modeling human health parameters. Toward that goal, novel features are proposed for wearable sensor modalities, namely wrist-worn motion sensors (accelerometer and gyroscope) and chest-worn electrocardiogram (ECG). These features are used in standard instance-based ML methods and sequential models in learning health parameters. To capture the temporal progression of health symptoms, sequential models and temporal ensemble of instance-based models are designed and compared. To overcome the challenges of noisy sparse labels from real-world data, multi-instance learning (MIL) methods are implemented to release the constraint of exact labeling of the whole sequence. Finally, to achieve improved performance and generalizability, this work proposes a novel contextual ensemble learning method, called ConxEns. This method leverages available contextual information in learning ‘weak’ contextual models, and implement a probabilistic aggregator to infer the health parameters as an ensemble of the inferences of the specialized models. ConxEns is implemented for both classification and regression task and is evaluated for performance improvement, and to demonstrate its potential in generalizing beyond known dataset.
The proposed models and ConxEns pipelines are evaluated in the scope of two major healthcare studies: dementia and asthma care. For the dementia study, the BESI system is evaluated and used to collect patients’ wrist motion signals during month-long deployments at their homes. The agitated behavior in dementia patients are modeled as a classification task using the proposed motion features with standard ML models. Multi-instance models MIL-Boost and MI-SVM have demonstrated improved performance compared to single-instance models. Using agitated behavioral symptoms as a context, the proposed ConxEns pipeline is implemented to predict agitation. Result demonstrates robust performance and generalizability. The proposed solutions are applied to another study on asthma care demonstrating the generalizability across health applications. For this study, novel features are proposed for wrist motion and wearable ECG signals and are used to regress the respiratory parameters of study participants using both standard ML models and the proposed ConxEns pipeline. Improved performance and generalization in evaluation demonstrate the robustness of the proposed method for health monitoring. This work leads the way for future research in personalized health modeling toward explainable intervention design.
