Abstract
Advances in integrated circuit technologies have enabled engineers to embed computation and communication resources in small form factors. Small wireless sensors have enabled a variety of ambulatory medical detection applications where individuals are instrumented in order to detect events outside a clinical setting. To make such devices suitable for long-term use, small batteries must be used. Thus, reducing overall energy consumption is essential.
In these applications, reducing the amount of data sampled or processed can reduce energy consumption. One way to do this is to use a low-cost screening detector that quickly rules out segments of the sampled data that are unlikely to contain an event. Depending on the detection ability and relative power consumption of the screening detector, the combination of the original detector and screening detector can decrease energy consumption dramatically while maintaining low false positives and false negatives.
In this paper, we describe a systematic method that uses machine learning to construct a screening detector for multi-feature detection algorithms. We evaluate this technique on real electroencephalography (EEG) data obtained from patients with epileptic seizures. Our results suggest that for most patients our technique can be used to construct a detector that reduces computation time by an average of 80% relative to the original detector. Moreover, we estimate that the average reduction in energy consumption would be 71% overall and 77% during non-seizure periods. This corresponds to a 4x increase in battery lifetime.
© ACM, 2008. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in HealthNet '08: Proceedings of 2nd International Workshop on Systems and Networking Support for Health Care and Assisted Living Environment, http://doi.acm.org/10.1145/1515747.1515767
[ paper (pdf) | presentation (pdf) | presentation (ppt) ]
Update: Since the paper was published, I have made a few spelling corrections and added a table. You can download the latest version of the paper here.
If you choose to use slides from my presentation, I would appreciate if you would include proper credit. Thanks.
Software and Data
Most of the code for this work was written in MATLAB. I will try to release the code shortly. Unfortunately, the data for this work is not available at this time due to privacy constraints. The EEG data was collected from real patients and an IRB needs to be filed in order for us to release the data.
Back to Publications