Samenvatting
Aim: The aim of this study is to explore the possibility to use EMG based features to build an alarm system to detect tonic epileptic seizures during sleep. It is known that patients suffering from severe motor seizures have the highest risk on SUDEP. A sufficient alarm system might reduce the SUDEP rate.
Methods: Data of three epilepsy patients with tonic seizures were used in this project. EMG is recorded of the Deltoid muscle on the left and right side with a sample frequency of 200 Hz. After preprocessing, five features are calculated for every second in the data: root mean square (RMS), wavelength (WL), median frequency (MDF), zero crossing rate (ZCR) and the coherence.
An algorithm, based on data of one patient (trainingset), is created to generate an alarm when a feature exceeds a certain threshold. ROC curves are used to determine the optimal threshold values for every feature based on a minimum sensitivity of 0.8. The optimal features are defined as the ones with the best combination of statistical measures (sensitivity, specificity and positive predictive value(PPV)). The WL and the ZCR turned out to be the most feasible features for the purpose of tonic seizure detection during the night.
Results of the designed algorithm: Tonic seizure detection, based on only WL, results in a sensitivity of 0.86 in the trainingset. By adding a second feature, ZCR, the sensitivity can be increased to 1. However, the specificity and the PPV decrease indicating an increase in false positives. Using WL and ZCR, the sensitivity for the three datasets were 0.92, 0,09 and 1 and the PPV values were 0.73. 1 and 0.56.
Discussion: The results of dataset 1 and 3 are close to optimal, in contrast to the results of dataset 2. Dataset 2 contains multiple seizure types and the tonic seizures are not as pronounced compared to the other datasets. Dataset 1 and 3 represent data from the same patient. No normalization has been applied in this project so the threshold values might not be applicable for seizure detection in dataset 2.
Conclusion and future recommendations: Based on EMG signals, sampled with a frequency of 200 Hz, it is not possible to generate a reliable alarm system to detect tonic seizures during the night. To improve the algorithm, it might be necessary to use a better trainingset representing seizures from a larger patient population or to use patient specific algorithms. The addition of different modalities, for example ECG or ACM, might improve the results of the algorithm further.
Methods: Data of three epilepsy patients with tonic seizures were used in this project. EMG is recorded of the Deltoid muscle on the left and right side with a sample frequency of 200 Hz. After preprocessing, five features are calculated for every second in the data: root mean square (RMS), wavelength (WL), median frequency (MDF), zero crossing rate (ZCR) and the coherence.
An algorithm, based on data of one patient (trainingset), is created to generate an alarm when a feature exceeds a certain threshold. ROC curves are used to determine the optimal threshold values for every feature based on a minimum sensitivity of 0.8. The optimal features are defined as the ones with the best combination of statistical measures (sensitivity, specificity and positive predictive value(PPV)). The WL and the ZCR turned out to be the most feasible features for the purpose of tonic seizure detection during the night.
Results of the designed algorithm: Tonic seizure detection, based on only WL, results in a sensitivity of 0.86 in the trainingset. By adding a second feature, ZCR, the sensitivity can be increased to 1. However, the specificity and the PPV decrease indicating an increase in false positives. Using WL and ZCR, the sensitivity for the three datasets were 0.92, 0,09 and 1 and the PPV values were 0.73. 1 and 0.56.
Discussion: The results of dataset 1 and 3 are close to optimal, in contrast to the results of dataset 2. Dataset 2 contains multiple seizure types and the tonic seizures are not as pronounced compared to the other datasets. Dataset 1 and 3 represent data from the same patient. No normalization has been applied in this project so the threshold values might not be applicable for seizure detection in dataset 2.
Conclusion and future recommendations: Based on EMG signals, sampled with a frequency of 200 Hz, it is not possible to generate a reliable alarm system to detect tonic seizures during the night. To improve the algorithm, it might be necessary to use a better trainingset representing seizures from a larger patient population or to use patient specific algorithms. The addition of different modalities, for example ECG or ACM, might improve the results of the algorithm further.
Originele taal2  Engels 

Begeleider(s)/adviseur 

Datum van toekenning  31 okt 2014 
Plaats van publicatie  Eindhoven 
Uitgever  
Status  Gepubliceerd  2014 