The objective of this project is to overcome the limitations of sensor artifacts (noise), false detection, and energy/power constraints by combining the analysis of multiple physiological signals through specialized hardware which implements a multi-layer classification technique comprised of a unique sequence of signal processing and machine learning functions to distill time series data. The hypothesis is that a hybrid architecture can leverage common operations and communication patterns between DSP and machine learning to support these computations more efficiently than traditional digital signal processors and general purpose processors. This hypothesis is explored in the context of wearable seizure detection, using traces of EEG and other physiological sensor data obtained from the Epilepsy Center at University of Maryland Medical Center.
Success of this exploratory research could have a significant impact for robust and efficient monitoring and use of continuous multi-physiological data for patients. Just in the context of epilepsy, it could enable seizure detection and caregiver alerts, which is important at night when seizures can happen without someone to help nearby. Longer term potential impacts extend to human-centered cyber-physical systems, cyber-security, and unmanned vehicles.
