This study evaluated whether expert-confirmed four-second interictal epileptiform discharge (IED) epochs can be discriminated among five predefined scalp-distribution categories using handcrafted features, multiple classifiers, channel ablation, and SHAP explainability. A total of 2,514 expert-confirmed epochs labeled as generalized, frontal, temporal, occipital, or centro-parietal were analyzed. Identical stratified epoch-level partitions, training-only SMOTE, and 26 handcrafted features per included channel were applied across channel configurations. Linear discriminant analysis achieved the best EEG-only test accuracy (88.89%). Adding two ECG channels increased descriptive test accuracy across classifiers; CatBoost reached 93.25% with EEG+ECG and 94.44% with the full 29-channel input. SHAP analysis highlighted ECG channel contributions and identified beta-band power as the most important feature family in the CatBoost EEG+ECG model. The analysis is explicitly framed as conditional spatial classification after expert IED confirmation and does not claim physiological localization or clinical biomarker validity.
Epilepsy evaluation commonly relies on EEG to detect and describe interictal epileptiform discharges. Beyond detecting whether an IED is present, clinicians often assign a scalp-distribution label describing the electrodes showing maximal activity. The scalp-distribution label used here—generalized, frontal, temporal, occipital, or centro-parietal—reflects surface manifestation, not source localization or seizure-onset zone.
This work treats scalp-distribution assignment as a distinct downstream classification task that begins after expert confirmation of an IED. By conditioning on epochs already labeled as containing an IED, the study isolates the spatial categorization problem from the separate detection problem and enables controlled comparisons of channel-set contributions and model behaviour.
The study set five goals: (1) test separability of the five expert-defined scalp-distribution categories using a common handcrafted-feature representation; (2) compare multiple machine-learning classifiers under identical epoch-level partitions; (3) quantify descriptive performance change when two ECG channels are added to 19-channel scalp EEG; (4) compare EEG with ECG against a complete 29-channel input that also includes referential and EMG channels; and (5) apply SHAP to characterize model dependence on specific channels and features while avoiding physiological or clinical localization claims.
The analysis used 2,514 expert-confirmed four-second IED epochs drawn from a publicly available dataset referenced by the authors. Each epoch had been labeled by experts into one of five scalp-distribution categories: generalized, frontal, temporal, occipital, or centro-parietal. The dataset includes scalp EEG channels plus referential, ECG, and EMG channels; the channel-ablation design exploits this multimodal recording to test how auxiliary channels affect downstream spatial categorization.
For every included channel, 26 handcrafted features spanning time-domain, frequency-domain, and nonlinear measures were extracted independently. The feature set produced channel-wise descriptors that form the input to classifiers. The authors note that feature attributions computed later by SHAP are conditional on this representation and on the fitted model.
Three primary channel configurations were used so that ablation measured the incremental contribution of added channel groups under identical preprocessing and modelling:
This staged comparison isolates descriptive effects of adding ECG and then the additional referential/EMG channels. The design does not attribute any performance change to specific physiology; rather, it tests whether classifiers exploit information outside scalp EEG.
All experiments used identical stratified epoch-level train/validation/test partitions. Class imbalance was addressed using SMOTE applied only to the training folds. Multiple classifier families were trained on the same partitions and feature matrices to assess consistency of channel-set effects across modelling assumptions.
SHapley Additive exPlanations (SHAP) were used to identify which channel-feature variables influenced fitted models. SHAP attributions were interpreted as model-specific contributions to predictions and not as evidence of causal physiological mechanisms or validated biomarkers. The combination of ablation and SHAP was intended to reveal whether observed performance gains coincided with substantive model dependence on added modalities.
The controlled conditional design isolates spatial categorization among expert-labeled IED epochs but does not address IED detection, source localization, or generalization to unseen patients. The reported SHAP attributions indicate which channel-feature inputs the fitted models used for prediction in this dataset and representation; they should not be interpreted as physiological localization or validated ECG biomarkers. The increased accuracy after adding ECG is descriptive only, since formal paired significance testing was not performed. Other limitations—such as potential dataset-specific confounds, patient-state dependence, or contamination effects—are acknowledged by the authors.
Using a common handcrafted-feature representation and identical partitions, the study demonstrates that expert-confirmed IED epochs can be conditionally classified among five scalp-distribution categories with high descriptive accuracy. Staged channel ablation showed descriptive gains when ECG and additional channels were included, and SHAP identified ECG channels and beta-band power as prominent contributors to the fitted models. The results provide a controlled benchmark for downstream annotation and model interpretability while explicitly avoiding claims about physiological localization, unseen-patient generalization, or clinical biomarker status.