Multi-sensor data fusion for ambulatory health monitoring: signal processing and deep learning techniques

With the increasing popularity of wearable IoT devices, there is a growing demand for health-related parameter estimation using wearable sensor signal processing. This is further exacerbated by the increasing cost of healthcare, at a rate of 4-6\% over the past several years, outstripping average economic growth. Continuous monitoring of physiological signals using wearable Internet of Things (IoT) sensors for early detection and preventative action is widely considered a solution to tackle the costs and risks associated with these issues. An important challenge in wearable health monitoring is the low and unpredictable quality of the signals obtained from the wearable devices due to noises such as motion artifacts, muscle noise, etc. Fusing data from multiple sensors has emerged as an efficient solution to overcome the challenge of low signal quality, occasional loss in information, or spurious data. This is commonly known as multi-sensor fusion wherein information obtained from multiple sensors is fused to achieve improved data interpretation accuracy over that of using a single sensor source. In this thesis, various data fusion techniques are developed for wearable sensing applications, with the goal of proposing a generalizable fusion framework for wearable health monitoring. This thesis explores a serialized approach to developing a generalized fusion framework. This is achieved by developing multi-sensor data fusion algorithms for signals of the same modalities in the initial stage. This fusion approach is applied to the heart rate estimation problem through the fusion of electrocardiogram (ECG) signals. The heart rate estimation fusion approach is further generalized to the fusion of multi-modal signals for temporal inferencing. To carry out the fusion process based on the real-time quality of the acquired signals, a metric called signal quality indicator is developed which can be used for quality assessment of any quasi-periodic signal. The proposed fusion architecture was applied to the heartbeat detection problem using ECG and photoplethsymogram (PPG) signals to showcase the merits of the fusion architecture. Further, a fusion framework for signals sampled at different sampling frequencies for temporal inferences was developed. The advantage of this framework is that no padding/ resampling process, which can increase computational costs, was used to aid the fusion process. The computational complexity of the proposed framework was studied to analyze the suitability of its deployment in wearable IoT sensors. The developed fusion framework was applied to the sleep apnea detection problem by fusing ECG, peripheral oxygen saturation (SpO2), and abdominal movement signals. Finally, a fusion framework is proposed, wherein the fusion of multi-modal data can be carried out at variable information abstraction levels which is automatically learned by the fusion algorithm, depending on the application at hand. This framework was applied to the atrial fibrillation detection problem through the fusion of ECG and PPG signals. This fusion framework can be considered as the generalized version of the fusion frameworks already developed in this thesis. The research presented in this thesis thus establishes the potential of data fusion methodologies for improved inferences in wearable physiological signal monitoring devices and the fusion frameworks developed can be applied to other analogous classes of problems.