Optimal Metamodeling to Interpret Activity-Based Health Sensor Data

Wearable sensors are revolutionizing the health monitoring and medical diagnostics arena. Algorithms and software platforms that can convert the sensor data streams into useful/actionable knowledge are central to this emerging domain, with machine learning and signal processing tools dominating this space. While serving important ends, these tools are not designed to provide functional relationships between vital signs and measures of physical activity. This paper investigates the application of the metamodeling paradigm to health data to unearth important relationships between vital signs and physical activity. To this end, we leverage neural networks and a recently developed metamodeling framework that automatically selects and trains the metamodel that best represents the data set. A publicly available data set is used that provides the ECG data and the IMU data from three sensors (ankle/arm/chest) for ten volunteers, each performing various activities over one-minute time periods. We consider three activities, namely running, climbing stairs, and the baseline resting activity. For the following three extracted ECG features – heart rate, QRS time, and QR ratio in each heartbeat period – models with median error of <25% are obtained. Fourier amplitude sensitivity testing, facilitated by the metamodels, provides further important insights into the impact of the different physical activity parameters on the ECG features, and the variation across the ten volunteers.