Conditional spatial classification of expert-confirmed interictal epileptiform discharge epochs: An EEG-ECG ablation and SHAP analysis.

Adding ECG channels improved internal classification of where already confirmed epileptiform discharges appeared on the scalp. The model did not detect discharges, localize a seizure source or prove that ECG is a clinical biomarker.

Plabon AM, Mukit A, Neyamul M, Jehady OF, Zuba FT, Mina MF, Islam T.
PloS one · Publication date: 2026-09-22
Added to site: 2026-09-26
DOI: 10.1371/journal.pone.0358782

Electronic publication date verified in Europe PMC and the full text; first indexed 2026-09-23. This source date is separate from the site's addition date.

Modalities

EEG, ECG, Machine learning

Clinical focus

Epilepsy, Interictal epileptiform discharges, Signal classification

Population

Clinical EEG recordings from the Epilepsy Center of Peking Union Medical College Hospital

Sample

2,514 expert-confirmed four-second IED epochs from 52 participant recordings containing IEDs within an 84-participant dataset

Intervention / method

Five-class spatial categorization of already confirmed IED epochs using 19 scalp EEG channels, EEG plus two ECG channels or the full 29-channel EEG, referential, ECG and EMG input

Comparator

Channel-set ablation across EEG only, EEG plus ECG and all 29 channels using identical epoch-level partitions

Clinician summary

Best test accuracy was 88.89% with EEG only, 93.25% with EEG plus ECG and 94.44% with all channels. Every directly comparable classifier improved numerically after ECG was added, but the changes were descriptive and not tested with paired statistical comparisons.

How to interpret the findings

The split was epoch-level rather than patient-disjoint, allowing participant-specific information to enter training and test sets. Only expert-confirmed IED epochs were included, so specificity, false-positive reduction and screening were not evaluated. Prior methods were not reimplemented, and the full-channel condition jointly added referential and EMG channels.

Equipment and disclosures

Nineteen international 10-20 scalp EEG channels, four referential channels, two ECG channels and four deltoid EMG channels were available; acquisition hardware was not reported in the checked full text.

The full text states that the authors received no specific funding and declared no competing interests.

Related evidence

Compare findings from the same clinical area, including mixed results.

• Interictal electric source imaging with open-source software and routine low-density electroencephalography: A systematic comparison of four inverse methods in drug-resistant epilepsy. · Positive with limitations
• Bridging Temporal and Spatial Scales: Contemporary Advances in EEG-fMRI Integration With a Focus on Epilepsy. · Positive with limitations
• Predicting levetiracetam neuropsychiatric adverse effects using graph-theoretical electroencephalographic network metrics in drug-naive focal epilepsy: A multicenter longitudinal study. · Positive with limitations
• ADHD classification using spherical phase space partitioning and symbolic time series analysis of multi-channel EEG signals · Positive with limitations