Detection of epileptic seizures by pattern recognition from wireless accelerometer data

J. Stigwall, A. Hildeman, J. Wipenmyr, T. Pettersson, K. Malmgren, B. Rydenhag
Type of publication: 
Conference item

1. Purpose
Many persons with epilepsy develop drug-resistance and require continued monitoring for treatment evaluation. To help physicians revise antiepileptic drugs quicker and more accurately than manual seizure diaries allow, we have developed a system for automatic logging of seizures based on custom wireless inertial motion sensors.  
2. Method
Motion data has been recorded using wireless acceleration sensors mounted on three locations on persons with epilepsy as they have been admitted to Sahlgrenska University Hospital for EEG monitoring. The dataset includes 38 patients and covers over 130 days and 200 epileptic seizures. EEG recordings have been used as reference to provide accurate seizure timing information.
3. Results
The wireless sensors have been well received, with very few patients experiencing discomfort. Different approaches to the machine learning problem have been investigated, resulting in a signal processing system which uses both short-term seizure-unique motion features as well as longer term seizure progression information.  On patients with generalized tonic-clonic seizures it is usual to see 100% sensitivity with no false positives, and for patients with weaker motoric seizures the sensitivity typically reaches 90%-95% with only 15-20% false positives (relative to the total number of positives).
4. Conclusion
Automatic logging of epileptic seizure based solely on motion data has shown to be a reliable technique compared to manual seizure diaries. Using small wireless sensors that can be mounted e.g. in clothing it also provides a much less obtrusive alternative than systems that rely on EEG data.

Published in: 
Proceeding of 29th International Epilepsy Congress, Rome, Italy