EEG-based neonatal seizure detection with Support Vector Machines

Abstract

Objective: The study presents a multi-channel patient-independent neonatal seizure detection system based on the Support Vector Machine (SVM) classifier.

Methods: A machine learning algorithm (SVM) is used as a classifier to discriminate between seizure and non-seizure EEG epochs. Two post-processing steps are proposed to increase both the temporal precision and the robustness of the system. The resulting system is validated on a large clinical dataset of 267 h of EEG data from 17 full-term newborns with seizures.

Results: The performance of the system using event-based metrics is reported. The system showed the best up-to-date performance of a neonatal seizure detection system. The system was able to achieve an average good detection rate of ~89% with one false seizure detection per hour, ~96% with two false detections per hour, or ~100% with four false detections per hour. An analysis of errors revealed sources of misclassification in terms of both missed seizures and false detections.

Conclusions: The results obtained with the proposed SVM-based seizure detection system allow for its practical application in neonatal intensive care units.

Significance: The proposed SVM-based seizure detection system can greatly assist clinical staff, in a neonatal intensive care unit, to interpret the EEG. The system allows control of the final decision by choosing different confidence levels which makes it flexible for clinical needs. The obtained results may provide a reference for future seizure detection systems.

Fig. A1

Two-class linear classification. The support vectors are indicated with crosses.

Fig. 1

(a) Architecture of the SVM-based seizure detection system. Various stages of the algorithm such as feature extraction, classification and post-processing are schematically shown. (b) The detailed structure of SVM classification which includes converting raw SVM outputs to probabilistic values.

Fig. 2

Effects of the post-processing scheme. (a) The raw output of the SVM classifier. (b) The output converted to a probability via a sigmoid function. (c) The smoothed output after a 15-tap moving average filter is applied. (d) The binary decisions resulting from applying a threshold of 0.5 to the filtered probabilities of seizure. (d) All seizures are detected using binary decisions however total seizure burden is underestimated. (e) The binary decisions from another EEG channel where one seizure is missed. (f) The binary decisions after two channels which are shown in plot (d) and (e) are fused. (g) The final binary decisions after the collar operation, which increases the duration of all positive decisions. (h) The ground truth, where 1 indicates seizure.

Fig. 3

The curve of variation of the GDR against FD/h.

Fig. 4

Detection of seizures of different duration by the SVM-based seizure detection system at 1 FD/h. The number of seizures in each time category is 72, 182, 240, and 197.

Fig. 5

Example of a missed seizure. The seizure is indicated with arrows in channels C3–O1 and C3–T3.

Fig. 6

Most frequent sources of false detections produced by the SVM-based seizure detection system at 1 FD/h.

Fig. 7

Example of background activity (trace-alternant) incorrectly classified as a seizure in channel F3–C3.

Fig. 8

Example of an electrode-disconnect artifact after filtering incorrectly classified as a seizure in channels F3–C3 and Cz–C3.

Fig. 9

Example of a respiratory artifact incorrectly classified as a seizure in channel T4–C4.

Fig. 10

Example of a “being handled” artifact incorrectly classified as a seizure in channel C3–T3.

Fig. 11

Example of a short duration seizure that was not annotated by neurophysiologists but detected with high confidence by the algorithm in channel F4–C4.