Online Prediction of Lead Seizures from iEEG Data

Abstract

We describe a novel system for online prediction of lead seizures from long-term intracranial electroencephalogram (iEEG) recordings for canines with naturally occurring epilepsy. This study adopts new specification of lead seizures, reflecting strong clustering of seizures in observed data. This clustering results in fewer lead seizures (~7 lead seizures per dog), and hence new challenges for online seizure prediction, that are addressed in the proposed system. In particular, the machine learning part of the system is implemented using the group learning method suitable for modeling sparse and noisy seizure data. In addition, several modifications for the proposed system are introduced to cope with the non-stationarity of a noisy iEEG signal. They include: (1) periodic retraining of the SVM classifier using most recent training data; (2) removing samples with noisy labels from training data; and (3) introducing a new adaptive post-processing technique for combining many predictions made for 20 s windows into a single prediction for a 4 h segment. Application of the proposed system requires only two lead seizures for training the initial model, and results in high prediction performance for all four dogs (with mean 0.84 sensitivity, 0.27 time-in-warning, and 0.78 false-positive rate per day). The proposed system achieves accurate prediction of lead seizures during long-term test periods, 3-16 lead seizures during a 169-364 day test period, whereas earlier studies did not differentiate between lead vs. non-lead seizures and used much shorter test periods (~few days long).

Keywords: group learning; iEEG; lead seizure; non-stationarity; seizure prediction; support vector machines; unbalanced classification.

Figure A1

Clustering of seizures for canine data (L7): Explanation: among all 105 seizures (shown in red circles), 103 seizures occur in clusters, 2 seizures are isolated. Using T = 3 days, 8 clusters (~lead seizures) can be detected. Cluster duration is between 1 to 2 days, duration between two lead seizures is between 6 and 304 days.

Figure A2

Clustering of seizures for canine data (M3): Explanation: among all 29 seizures (shown in red circles), 11 seizures occur in clusters, 18 seizures are isolated. Using T = 3 days, 22 clusters (~lead seizures) can be detected. Cluster duration is between 1 to 4 days, duration between two lead seizures is between 5 and 43 days.

Figure A3

Clustering of seizures for canine data (P2): Explanation: among all 22 seizures (shown in red circles), All 22 seizures occur in clusters. Using T = 3 days, 5 clusters (~lead seizures) can be detected. Cluster duration is between 0 to 1 days, duration between two lead seizures is between 8 and 75 days.

Figure 1

Lead seizures of seizure clusters. Fourteen seizures are annotated in the recording period (represented as 14 red circles); Using T = 3 days, two of them are defined as lead seizures (represented as filled red circles).

Figure 2

Clustering of seizures for canine data (L2): (a) distribution of all 45 seizures (in red circles) during the whole period (July 2009–November 2010), indicating strong clustering. (b) distribution of 7 seizures (in blue circles) within one cluster highlighted in (a), during time period 3–5 June 2010.

Figure 3

System design parameters for online prediction: window, prediction period and prediction horizon.

Figure 4

Overview of the proposed system: (a) feature encoding for 20 s windows; (b) preprocessing (~window selection); (c) model estimation (~SVM training); (d) test/prediction stage.

Figure 5

Retraining procedure: The classifier is retrained every week, and predictions are made for the next 7-day period (shown in green color). Vertical red lines indicate lead seizures. Training data contain most recent interictal segments (shown in shaded region). Initial Training pool 1 contains 2 preictal segments and 16 interictal segments. Training pool 2 is extended to include fresh interictal segments (for training). Training pool N contains 6 preictal segments, but only 5 recent ones are used for training.

Figure 6

Histograms for 4 h iEEG segments and the corresponding bivariate decision space for discriminating between interictal (red lines/dots) and preictal (blue lines/dots) segments. Dashed lines denote thresholds, and green-shaded regions corresponds to preictal class segments. (a) Histograms for training data for Dog P2 (2 preictal and 16 interictal segments). (b) Histograms for test data for Dog P2 (1 preictal and 19 interictal segments). (c) Histograms for training data for Dog M3 (5 preictal and 40 interictal segments). (d) Histograms for test data for Dog M3 (1 preictal and 30 interictal segments).

Figure 7

Lead seizures (defined by T = 4 h) for dog L2 during observation period (July 2009–November 2010). Each circle represents an annotated seizure; shaded circles indicate 40 lead seizures (out of 45 total).

Figure 8

Lead seizures (defined by T = 3 day) for dog L2 during observation period (July 2009–November 2010). Each circle represents an annotated seizure; shaded circles indicate 8 lead seizures (out of 45 total).