CCF: SHF: Small: Collaborative Research: Domain-specific Reconfigurable Processor for Time-Series Data Mining and Monitoring

University Of New Mexico, Albuquerque NM

Investigators

Abstract

This proposal will investigate techniques to improve performance and reduce the cost and energy consumption of wearable devices (e.g., Internet-of-Things) that perform real-time medical monitoring. The objective is to demonstrate how to construct programmable integrated circuits that can provide monitoring capabilities while remaining small enough to be convenient and unobtrusive to the wearer. As a motivating example, such a device could detect and predict medical problems of fundamental importance such as pericardial tamponade, a life-threatening emergency where in the pericardium (a sac surrounding the heart) fills with fluid and prevents the heart from pumping blood, leading quickly and directly to death. Broader impacts of this effort include: reducing the cost of medical monitoring and saving lives; introduction of undergraduate students at both institutions to hardware/software co-design for wearable computing through a Freshman Discovery Seminar; inclusion of women and underrepresented minorities in the project; public release of hardware and software developed in the course of this project; and tutorial dissemination targeting the database/data mining and design automation research communities. The technical approach will involve the creation of application-specific integrated circuit hardware that can be added to embedded processors and microcontrollers to improve the performance of real-time medical monitoring applications. The research is based on the observation that real world data sets often exhibit a significant disparity in the dimensionality (data sampling rate) and cardinality (number of distinct values) of the data; for example, echocardiograms (ECGs) are often sampled at 1,024 Hz and 64-bits, although reducing the dimensionality to 128 Hz and the cardinality to 8-bits suffice for real-time monitoring. The fundamental challenge is that the minimum dimensionality and cardinality may vary from task-to-task, from individual-to-individual, and possibly even from hour-to-hour for a given individual. The Minimum Description Length (MDL) principle will be investigated as a potential solution to find the intrinsic dimensionality and cardinality of the data source, which can reduce data volume and improve detection accuracy by noise reduction. This information will then be leveraged to design domain-specific adaptive architectures that can exploit this reduced data volume to improve throughput and enhance battery lifetime.