An Empirical Investigation of Reconstruction-Based Models for Seizure Prediction from ECG Signals

TL;DR

This study tackles the challenge of predicting epileptic seizures from ECG signals as a practical alternative to EEG. It presents a reconstruction-based anomaly-detection framework that leverages time-frequency ECG representations (DWT, CWT, STFT) and three deep learning architectures—LSTM Autoencoder, Multi-Head Convolutional LSTM Autoencoder, and Transformer-based Anomaly Detection—trained in a patient-specific, unsupervised manner. Through moving-average smoothing and adaptive thresholding (\tau = \mu + k\sigma, with k = 2), the method achieves strong specificity (99.16%), competitive accuracy (76.05%), and a low false-positive rate (0.01/h), with an average lead time of about 40 minutes before seizure onset on the Siena database. The findings support ECG-based seizure prediction as a patient-friendly, near real-time solution and highlight the value of time-frequency features and hybrid architectures for capturing pre-ictal cardiac dynamics, while also pointing to the need for prospective validation and generalization across diverse cohorts.

Abstract

Epileptic seizures are sudden neurological disorders characterized by abnormal, excessive neuronal activity in the brain, which is often associated with changes in cardiovascular activity. These disruptions can pose significant physical and psychological challenges for patients. Therefore, accurate seizure prediction can help mitigate these risks by enabling timely interventions, ultimately improving patients' quality of life. Traditionally, EEG signals have been the primary standard for seizure prediction due to their precision in capturing brain activity. However, their high cost, susceptibility to noise, and logistical constraints limit their practicality, restricting their use to clinical settings. In order to overcome these limitations, this study focuses on leveraging ECG signals as an alternative for seizure prediction. In this paper, we present a novel method for predicting seizures based on detecting anomalies in ECG signals during their reconstruction. By extracting time-frequency features and leveraging various advanced deep learning architectures, the proposed method identifies deviations in heart rate dynamics associated with seizure onset. The proposed approach was evaluated using the Siena database and could achieve specificity of 99.16\%, accuracy of 76.05\%, and false positive rate (FPR) of 0.01/h, with an average prediction time of 45 minutes before seizure onset. These results highlight the potential of ECG-based seizure prediction as a patient-friendly alternative to traditional EEG-based methods.

Figures (7)

• Figure 1: Overall components of the proposed approach for Epileptic Seizure Prediction.
• Figure 2: Schematic Diagram of LSTM-AE. Enc is Encoder; Dec is Decoder; BN is Batch Normalization; DO is Dropout; LSTM is Long Short-Term Memory; and LS is Latent Space.
• Figure 3: Schematic Diagram of Multi-Head-Conv-LSTM-AE. Enc is Encoder; Dec is Decoder; BN is Batch Normalization; DO is Dropout; CNN is Convolution Layer; LSTM is Long Short-Term Memory; MH is Multi-Head Attention Layer; and LS is Latent Space.
• Figure 4: Schematic Diagram of Transformer (Enc-Enc). EM is Embedding; T-Enc is Transformer Encoder; LN is Layer Normalization; DO is Dropout; and FF is Feed-Forward Layer.
• Figure 5: Comparison of Raw and Smoothed Reconstruction Loss (Patient PN13, 1s segmentation, Scalogram, LSTM-AE). In the plots, the blue curve represents the reconstruction error, the red horizontal line indicates the threshold, the red dashed lines mark the seizure onset points, and the turquoise dash-dotted lines, highlight the pre-ictal period.