Detección de ataques epilépticos a partir de señales fisiológicas
Mejía Mejía, Elisa | 2019
More than 60 million people suffer from epilepsy, a disease due to neuronal hypersynchronous discharges. Refractory epilepsy is defined when patients do not respond to antiepileptic medications. Therefore, it is necessary to find other methods to control and monitor epileptic seizures. Physiological signals are alternatives to do this. Electroencephalography (EEG) is considered the gold standard method for seizure identification. EEG has the disadvantage of time-limited ambulatory recording. On the other hand, the autonomic nervous system (ANS) exercises control over the heart rate evaluable by electrocardiography (ECG). Epileptic seizures exert an autonomous effect on heart rate variability (HRV) that is a measurable, continuous and non-invasive indicator
In this project, different systems of recognition of epileptic seizures were evaluated through ECG signals using characteristics of the HRV and morphological, statistical and frequency characteristics of the ECG. This, implemented in a device that records ECG in real time, could help patients keep better control of their lives and provide information to doctors for a proper diagnosis and follow up.
For the development of the project, ECG signals were collected from databases available in free repositories and from institutions providing health services. The use of these signals was endorsed by the Institutional Ethics Committee of the EIA University. These signals were processed and filtered according to different signals quality indexes (SQI), and 56 characteristics of the HRV and ECG were extracted of the entire signal. These characteristics were analyzed by means of statistical methods to choose which contributed to the detection of epileptic seizures. Different neural network classifiers, support vector machines (SVN) and nearest neighbor methods (k-NN) that recognize these patterns were designed and, finally, the validity of these systems was evaluated to differentiate seizures.
The SVN with polynomial function of third order obtained 82.8% ± 8.9% sensitivity and 80.78% ± 11.75% specificity. The neural network with three hidden layers obtained 79.8% ± 5.1% sensitivity and 74.1% ± 6.8% specificity. Finally, the k-NN method with k = 3 obtained 75.6% ± 12.6% sensitivity and 68.3% ± 6.4% specificity. These systems of identification of epileptic seizures can contribute to the implementation of new non-invasive technologies for diagnosis and monitoring of patients with epilepsy. The best performing system was the support vector machine.