Interpretable Multimodal EEG-ECG Framework for Neonatal Seizure Prediction and Patient-Adaptive Transfer Learning

The gist Neonates are highly susceptible to seizures, often leading to short or long-term neurological impairments, and this study proposes a patient-independent convolutional neural network-based model for early prediction of neonatal seizures by distinguishing between interictal and preictal states of the EEG The overall results demonstrate the model’s potential for minimally supervised deployment in neonatal intensive care units, enabling timely and reliable prediction of neonatal seizures, while demonstrating strong generalization capability across unseen subjects through transfer learning.

Model Features and Inputs

The proposed model utilizes a reduced-montage EEG for minimal clinical intrusiveness, which eliminates the need for extensive scalp coverage in neonates, reducing discomfort and simplifying clinical setup without degrading performance. The input to the model is a multichannel mel-frequency cepstral coefficient matrix derived from EEG and ECG signals. This MFCC computation process comprises five key steps, including segmenting the signal into overlapping time frames and applying a Hamming window to minimize spectral leakage. The resulting MFCC feature matrices had dimensions of (19, 20, 11), where 19 corresponds to the total number of EEG and ECG channels.

Model Architecture

The proposed model architecture comprises three primary components: a CNN encoder, a channel attention mechanism, and a classifier for seizure prediction. The input to the model is a multichannel MFCC matrix derived from EEG and ECG signals. The CNN encoder is composed of three sequential 2D convolutional blocks, each including a 2D convolutional layer followed by batch normalization, a ReLU activation, dropout, and max-pooling. To enhance the model’s ability to focus on the most informative feature channels, a squeeze-and-excitation (SE) attention mechanism is integrated. The SE block consists of three main stages: Squeeze, Excitation, and Recalibration. The final component is a fully connected classifier that produces a probability score reflecting the likelihood of the input segment corresponding to a preictal state.

Training and Evaluation

To assess robustness, 10-fold cross-validation was employed, using binary cross-entropy loss function and optimized with the Adam optimizer. To address class imbalance in the datasets, class weights were incorporated into the loss function. The model performance was monitored using metrics including accuracy, sensitivity, specificity, and F1-score. Furthermore, a leave-one-patient-out (LOPO) evaluation was conducted to examine generalization capability. To address high intersubject variability in EEG patterns during LOPO, a transfer learning approach was adopted where the model is initially trained on data from all available neonates except the target subject and then finetuned using a short-duration signal segment comprising both interictal and preictal samples from the held-out neonate.

Interpretability

To improve interpretability, SHapley Additive exPlanations (SHAP) was used to interpret the model. This process involves constructing a test set comprising the first six preictal segments from the beginning of the preictal period and randomly selecting six interictal segments from each neonate. The resulting channel-wise SHAP values for each neonate were obtained using a formula that calculates importance based on mean and standard deviation of the 3D SHAP matrix. These processed channel importances are mapped onto the scalp electrode map to provide an approximate localization of seizure focus.

Results Summary

The proposed model achieved an average accuracy of 97.52%, sensitivity of 98.31%, specificity of 96.39%, and F1-score of 97.95% on the Helsinki neonatal EEG dataset. The inclusion of ECG alongside EEG improved the F1-score by 1.42%, while the incorporation of an attention mechanism yielded an additional 0.5% improvement. The model demonstrated strong performance, achieving over 96% and 95% in accuracy, sensitivity, specificity, and F1-score on the Helsinki and Siena datasets respectively. The ability to achieve strong performance with finetuning on short-duration data is particularly relevant in clinical contexts.

The study concludes that a lightweight CNN-based model integrated with attention mechanisms can predict neonatal seizures 30 minutes before onset, enabling timely clinical interventions. By finetuning with only a few samples, our method can generalize well to new subjects, demonstrating its applicability in realworld clinical settings. Our approach incorporates explainable AI techniques, generating scalp plots based on model predictions to help clinicians understand and localize upcoming seizures, enabling timely interventions and improved neonatal care. The overall results provide a strong foundation for future studies and highlight the potential of AI-driven seizure prediction to improve neonatal outcomes.

The gist Neonates are highly susceptible to seizures, often leading to short or long-term neurological impairments, and this study proposes a patient-independent convolutional neural network-based model for early prediction of neonatal seizures by distinguishing between interictal and preictal states of the EEG.

1. The model can achieve accurate seizure prediction up to 30 minutes before onset by utilizing a reduced-montage EEG for minimal clinical intrusiveness, which simplifies setup while retaining performance.

2. The system can be made more robust and efficient by employing a lightweight, multi-modal, and interpretable model, with an attention mechanism for enhanced feature relevance, as the authors state this results in a model with only about 70k parameters.

3. The prediction accuracy can be further enhanced by integrating complementary physiological data; specifically, the inclusion of ECG signals improved the F1-score by 1.42%, and incorporating an attention mechanism yielded an additional 0.5% improvement.

4. The system's clinical utility can be significantly increased through enhanced transparency, as it incorporates SHapley Additive exPlanations (SHAP) as an explainable artificial intelligence method to interpret the model and provided localization of seizure focus using scalp plots.

5. The model's generalization capability across unseen subjects can be improved by utilizing a transfer learning approach, where the pre-trained model is finetuned using a short-duration signal segment comprising both interictal and preictal samples from the held-out neonate.