Reducing the computational demands of medical monitoring classifiers by examining less data

Abstract

Instrumenting patients with small, wearable sensors will enable physicians to continuously monitor patients outside the hospital. These devices can be used for real-time classification of the data they collect. For practical purposes, such devices must be comfortable and thus be powered by small batteries. Since classification algorithms often perform energy-intensive signal analysis, power management techniques are needed to achieve reasonable battery lifetimes. In this thesis, we describe software-based methods that reduce the computation, and thus, energy consumption of real-time medical monitoring algorithms by examining less data. Though discarding data can degrade classification performance, we show that the degradation can be small. We describe and evaluate data reduction methods based on duty cycling, sensor selection, and combinations of the two. Random duty cycling was applied to an online algorithm that performs risk assessment of patients with a recent acute coronary syndrome (ACS). We modified an existing algorithm that estimates the risk of cardiovascular death following ACS. By randomly discarding roughly 40% of the data, we reduced energy consumption by 40%. The percentage of patients who had a change in their risk classification was 3%. A sensor selection method was used to modify an existing machine learning based algorithm for constructing multi-channel, patient-specific, delay-sensitive seizure onset detectors.

(cont.) Using this method, we automatically generated detectors that used roughly 60% fewer channels than the original detector. The reduced channel detectors missed seven seizures out of 143 total seizures while the original detector missed four. The median detection latency increased slightly from 6.0 to 7.0 seconds, while the average false alarms per hour increased from 0.07 to 0.11. Finally, we investigated the impact of approaches that combine duty cycling with sensor selection on the energy consumption and detection performance of the seizure onset detection algorithm. In one approach, where we combined two reduced channel detectors to form a single detector, we reduced energy consumption by an additional 20% over the reduced channel detectors.