Interpretable Multimodal EEG-ECG Framework for Neonatal Seizure Prediction and Patient-Adaptive Transfer Learning
Authors: Sithmini Ranasingha, Agasthi Haputhanthri, Hansa Marasinghe, Kithmin Wickremasinghe, Nima Wickramasinghe, Jithangi Wanigasinghe, Chamira U. S. Edussooriya, Joshua P. Kulasingham
Organizations: Department of Electronic and Telecommunication Engineering, University of Moratuwa, Sri Lanka · Department of Electrical and Computer Engineering, University of British Columbia, Canada · Department of Biomedical Engineering, University of Melbourne, Australia · Department of Pediatrics, Faculty of Medicine, University of Colombo, Sri Lanka
Neonatal seizures carry substantial risks of acute brain injury and long-term neurodevelopmental impairment, motivating the need for timely identification of impending seizure activity. While continuous electroencephalogram (EEG) monitoring remains the clinical gold standard, reliable patient-independent seizure prediction in neonates remains largely under-explored. We propose an interpretable multimodal framework that combines EEG and electrocardiogram (ECG) signals for neonatal seizure prediction. Our method uses mel-frequency cepstral coefficient (MFCC) representations processed by a lightweight (approx. 70k-parameter) convolutional neural network with squeeze-and-excitation attention. The framework was evaluated on the Helsinki neonatal EEG-ECG dataset and the Siena scalp EEG-ECG dataset. Under strict leave-one-patient-out (LOPO) evaluation, patient-independent prediction proved challenging (mean accuracy 58.5%, F1-score 60.3%), highlighting substantial inter-subject variability in neonatal EEG. However, fine-tuning the pretrained model with only five minutes of patient-specific data per class improved performance to 95.87% accuracy and 95.85% F1-score, demonstrating the feasibility of rapid patient-adaptive transfer learning. SHapley Additive exPlanations (SHAP) identified informative EEG channels that showed partial spatial agreement with clinically reported seizure origins. These findings suggest that fully patient-independent neonatal seizure prediction remains challenging, but that patient-adaptive fine-tuning with a small amount of data may provide a practical pathway toward personalized neonatal seizure prediction systems.
Figures & tables
Fig. 1: EEG and ECG recordings across seizure-related brain states. The figure illustrates interictal, preictal, ictal, and postictal states, along with the time durations considered for the Helsinki dataset.
Fig. 2: Scalp electrode placement and EEG channel selection. Selected electrodes are shaded in blue. Red arrows denote the twelve channels from [ 16 ] , while blue arrows indicate the six additional channels used in this study. Each arrow depicts the electrode pair for each channel, with arrowheads indicating the direction of the difference.
Work
Approx. No. of Parameters
Dissanayake et al. [ 8 ]
128k
Dissanayake et al. [ 7 ]
289k
Zhang et al. [ 10 ]
2000k
Ours
70k
TABLE I: Comparison of model complexity with representative patient-independent seizure prediction frameworks.
Fig. 3: Architecture of the proposed deep learning model for seizure prediction. Raw EEG and ECG are preprocessed, segmented into 5 s windows, and converted into MFCC matrices. Features are extracted using 2D convolutional layers with batch normalization and ReLU (blue) and max pooling layers (green). An attention mechanism precedes the dense classifier, which produces a binary output: 1 for preictal and 0 for interictal.
Hyperparameter
Helsinki
Siena
Learning rate
0.0004
0.002
Batch size
256
128
Weight decay
0.005
0.0001
Dropout
0.3
0.0001
Positive class weight
0.52
0.51
Reduction factor
8
16
TABLE II: Hyperparameter settings used for model training.
Neonate ID
LOPO
N=24 (Int: 1 min, Pre: 1 min)
N=120 (Int: 5 min, Pre: 5 min)
Acc
Sen
Spec
F1-Score
Acc
Sen
Spec
F1-Score
Acc
Sen
Spec
F1-Score
11
71.92
65.65
75.90
64.37
95.57
93.10
97.12
94.19
95.66
93.33
97.04
94.12
13
56.83
36.20
77.17
42.83
83.45
86.78
80.17
83.89
91.90
91.00
92.79
91.76
15
62.09
62.13
61.98
70.39
70.04
64.66
86.21
76.40
100
100
100
100
21
65.71
66.25
64.33
73.43
70.98
75.00
60.31
78.97
99.48
99.67
99.33
99.74
40
46.38
49.07
19.45
57.26
74.19
72.70
95.83
84.05
99.11
99.04
100
99.50
TABLE III: Subject-wise LOPO and fine-tuned results for each neonate using N=24 and N=120 , corresponding to 1 and 5 min of preictal/interictal data, respectively. Metrics are reported as percentages (%).
Duration
Input
Performance Metrics
Acc (%)
Sen (%)
Spec (%)
F1-score (%)
30 min
No attention
EEG
95.25±0.08
96.99±0.28
92.55±0.58
96.10±0.07
EEG + ECG
96.96±0.05
97.12±0.20
96.74±0.15
97.46±0.05
Attention
EEG
96.23±0.20
96.65±0.18
95.58±0.29
96.86±0.17
EEG + ECG
97.52±0.04
98.31±0.21
96.39±0.33
97.95±0.03
15 min
No attention
EEG
97.52±0.15
97.62±0.07
97.45±0.21
97.22±0.17
TABLE V: Sample-level 10-fold ablation study on the Helsinki dataset. Values are reported as mean ± std across trials.
Fig. 4: Scalp Mapping of SHAP Values for Preictal EEG Samples in Neonates. The primary seizure location from the dataset metadata is indicated below each plot. Shaded electrodes highlight the nine electrodes used in this study, while lines between electrode pairs represent the 18 EEG channels constructed using the difference of the respective electrodes, as mentioned in Fig. 2 . Line intensity corresponds to SHAP-based importance, with darker lines indicating greater importance.
Information and Communication Engineering Department National Engineering School of Tunis Tunis El Manar University · REGIM-Lab, University of Sfax, National Engineering School of Sfax (ENIS), BP 1173, Sfax, 3038, Tunisia
School of Computing and Digital Technologies, Sheffield Hallam University, Sheffield and University of Strathclyde, Glasgow, United Kingdom. · School of Computing and Digital Technologies, Sheffield Hallam University, Sheffield, United Kingdom · Department of Cmputational Intelligence, SRM Institute of Science and Technology,Kattankulathur-603203, India. +2