ECG-Based Detection of Epileptic Seizures in Real-World Wearable Settings: Insights from the SeizeIT2 Dataset

Figures

• 

  Figure A1. Sensitivity by seizure type (strict evaluation, no SDW) for all three anomaly-detection models (TimeVQVAE-AD, Matrix Profile, and MADRID) and their across-model average. Bars are grouped by seizure type and sorted by average sensitivity; the number of seizures per type is shown below each label.

• 

  Figure A2. Sensitivity by seizure type using the extended Seizure Detection Window (SDW; −5 min to +3 min) for all three anomaly-detection models (TimeVQVAE-AD, Matrix Profile, and MADRID) and their across-model average. Bars are grouped by seizure type and sorted by average sensitivity; the number of seizures per type is shown below each label.

• 

  Figure 1. Overview of the proposed ECG-based seizure detection pipeline using the SeizeIT2 dataset. The overall workflow is divided into a training phase and a test phase. The dataset is first split at the patient level into a training set (sub-001–096) and a test set (sub-097–125). ECG recordings undergo data quality checks (ECG availability, and saturation detection) and preprocessing (bandpass filtering, normalization, and downsampling), resulting in preprocessed training and test sets. In the training phase, all three anomaly-detection models (Matrix Profile, MADRID, and TimeVQVAE-AD) are trained and evaluated across multiple configurations, followed by post-processing via temporal clustering. Based on training performance, the best model and post-processing configuration are selected. In the test phase, this selected configuration is applied to the held-out test patients, and seizure events are detected and evaluated. Performance metrics are reported separately for all patients and for the subgroup of responders exhibiting pronounced ictal heart-rate changes.

• 

  Figure 2. Overview of the Matrix Profile-based anomaly-detection workflow. From a preprocessed ECG signal, all z-normalized subsequences of length m are extracted and compared using the Matrix Profile, which assigns to each position the distance to its nearest neighbor subsequence. Large values mark discords (candidate anomalies) which are then filtered through percentile thresholding and an overlap criterion to remove redundant or weak candidates.

• 

  Figure 3. Overview of the MADRID anomaly-detection workflow. The algorithm first extracts subsequences for multiple lengths m, ranging from mmin to mmax in steps of Δm (step size). For each subsequence, MADRID computes a discord score that quantifies the degree of deviation from typical patterns in the time series; higher scores indicate stronger dissimilarity. The resulting multi-length discord matrix is refined through a three-stage anomaly-selection procedure: (1) Global percentile filtering: subsequences below a global percentile threshold are removed, ensuring that only strongly deviant patterns remain. (2) Per-length ranking and top-k selection: subsequences are ranked by discord score for each length and the top k candidates are retained. (3) Overlap constraint: candidates overlapping more than 25% with already selected anomalies are discarded, yielding a final set of non-redundant anomaly candidates. This filtered anomaly list forms the basis for post-processing, clustering, and seizure-detection evaluation.

• 

  Figure 4. TimeVQVAE-AD overview. During training (Stages 1 and 2), only normal intervals are used; anomalous regions are excluded. Stage 1 (Tokenizer): Normal segments are transformed by the STFT and passed through the encoder and vector quantizer to obtain the token map s∈{1,…,K}H×W. The encoder, vector quantizer, and decoder are optimized with a reconstruction objective; iSTFT of the decoded STFT yields an approximate waveform. Green boxes denote modules optimized in this stage. Stage 2 (Prior): The tokenizer is frozen. From the same normal-only data, randomly chosen windows M are masked and a bidirectional transformer prior is trained to predict masked tokens by maximizing pθ(s∣sM), equivalent to minimizing masked negative log-likelihood. Green boxes denote modules optimized in this stage. Evaluation: With the tokenizer and prior fixed, a masking window of width α slides along time. For each center index w, the prior outputs categorical distributions that quantify how likely the masked tokens are under the learned normal model; the anomaly score is the mean negative log-likelihood over the masked span. Repeating over w yields a score map a˜. Scores across multiple α may be summed, and a…

• 

  Figure 5. Illustration of the Seizure Detection Window (SDW) concept on a synthetic ECG example. The colored regions mark the ictal period and the extended SDW interval from five minutes before to three minutes after seizure onset. Anomalies occurring outside the SDW are treated as false positives, whereas anomalies within the SDW are counted as true positives, reflecting clinically meaningful pre- and post-ictal autonomic changes. (a) Without SDW: only ictal anomalies are counted as true positives. (b) With SDW: pre- and post-ictal anomalies inside the –5 min, +3 min window are also counted as true positives.

• 

  Figure 6. Average seizure-type-specific sensitivity across all models in the sensitivity-optimized configuration, evaluated under the strict detection criterion (no SDW). Sensitivity values represent the mean sensitivity across TimeVQVAE-AD, Matrix Profile, and MADRID; numbers below each bar indicate the number of seizures of that type in the test set.

• 

  Figure 7. Average seizure-type–specific sensitivity across all models in the sensitivity-optimized configuration, evaluated under the expanded detection criterion (SDW). Sensitivity values represent the mean sensitivity across TimeVQVAE-AD, Matrix Profile, and MADRID; numbers below each bar indicate the number of seizures of that type in the test set.

• 

  Figure 8. Comparison of EEG-based and ECG-based seizure detection performance on the SeizeIT2 dataset. (a) Sensitivity of ChronoNet and SVM (EEG-based baselines from Bhagubai et al. [11]) compared with TimeVQVAE-AD evaluated on ECG data in its sensitivity-optimized and false-alarm-optimized configurations. (b) Corresponding false-alarm rates (FA/h) for the same models. The figure highlights the fundamental trade-off between sensitivity and false-alarm rate and demonstrates that ECG-based TimeVQVAE-AD can achieve sensitivity comparable to ChronoNet while producing substantially fewer false alarms.