Predicting epileptic seizures in advance

Abstract

Epilepsy is the second most common neurological disorder, affecting 0.6-0.8% of the world's population. In this neurological disorder, abnormal activity of the brain causes seizures, the nature of which tend to be sudden. Antiepileptic Drugs (AEDs) are used as long-term therapeutic solutions that control the condition. Of those treated with AEDs, 35% become resistant to medication. The unpredictable nature of seizures poses risks for the individual with epilepsy. It is clearly desirable to find more effective ways of preventing seizures for such patients. The automatic detection of oncoming seizures, before their actual onset, can facilitate timely intervention and hence minimize these risks. In addition, advance prediction of seizures can enrich our understanding of the epileptic brain. In this study, drawing on the body of work behind automatic seizure detection and prediction from digitised Invasive Electroencephalography (EEG) data, a prediction algorithm, ASPPR (Advance Seizure Prediction via Pre-ictal Relabeling), is described. ASPPR facilitates the learning of predictive models targeted at recognizing patterns in EEG activity that are in a specific time window in advance of a seizure. It then exploits advanced machine learning coupled with the design and selection of appropriate features from EEG signals. Results, from evaluating ASPPR independently on 21 different patients, suggest that seizures for many patients can be predicted up to 20 minutes in advance of their onset. Compared to benchmark performance represented by a mean S1-Score (harmonic mean of Sensitivity and Specificity) of 90.6% for predicting seizure onset between 0 and 5 minutes in advance, ASPPR achieves mean S1-Scores of: 96.30% for prediction between 1 and 6 minutes in advance, 96.13% for prediction between 8 and 13 minutes in advance, 94.5% for prediction between 14 and 19 minutes in advance, and 94.2% for prediction between 20 and 25 minutes in advance.

Figure 1. Invasive EEG recording of patient 2 from the Freiburg EEG Database.

The image corresponds to pre-seizure and seizure data. Each row displays one of the 6 channel recordings. The name of the relevant EEG channel is listed to the right of each signal. The EEG signals were visualised using EEGLAB software .

Figure 2. Signal Energy over 6 EEG channels for patient 2 from the Freiburg EEG Database.

There is ictal activity from seconds 5 through 35. SE stands for Signal Energy.

Figure 3. Accumulated Energy over 6 EEG channels for patient 2 from the Freiburg EEG Database.

There is ictal activity from seconds 5 through 35. AE stands for Accumulated Energy.

Figure 4. An annotated epoch of the Invasive EEG of an epileptic seizure.

All four states of ictal, pre-ictal, ictal, post-ictal and inter-ictal are colour coded. EEG signals belong to patient 2 from the Freiburg EEG Database and were visualised using the EEGLAB software .

Figure 5. The architecture of ASPPR.

The system consists of a Pre-processing Module and a Learning Module. The data preparation and initial experimental setup takes place in the Pre-processing Module, which varies for each experiment. This is separated from the learning and classification task in the Learning Module.

Figure 6. Amendment of datasets for time-in-advance predictive models.

The top image displays a standard Ictal file where the ictal data is immediately preceded by a 300-seconds period of pre-ictal data, and also represents the dataset used to build ‘t = 0’ predictive models. The subsequent series of images then illustrates, from top to bottom, the datasets used to build the ‘t = 1’, ‘t = 2’, ‘t = 3’, ‘t = 4’ and ‘t = 5’ predictive models, constructed from the ‘t = 0’ model via removal of t minutes of immediately pre-ictal data, and relabeling (as pre-ictal) t minutes of inter-ictal data.

Figure 7. Summary results of advance prediction by ASPPR on 21 patients.

The plot shows Accuracy, Sensitivity, Specificity and S1-Score averaged across all 21 patients at each prediction time-step. The plot also displays the minimum, mean, maximum and full feature-set values for the S1-Score measure as well as the benchmark S1-Score value.

Figure 8. Distribution of S1-scores over individual patients for each time-in-advance prediction model.

The boxes at each interval display the distribution of average S1-Score of each of the 21 patients at each advance prediction time-step.

Figure 9. Mean S1-Score of the ASPPR algorithm for 21 patients.

The legend orders the patients in ascending order of their S1-Score averaged over all time-steps.