CHB-MIT Pediatric Seizure Detection Classifier

What this research found

Published seizure-detection accuracies on pediatric scalp electroencephalography are often inflated by training on the same child the model is later tested on, or by letting correlated windows from one recording straddle the train-test boundary. K-Dense built a deliberately leakage-free alternative: a random forest on 180 interpretable features, evaluated so that no window from a held-out child ever appears in training. Pooled across 49,959 windows it reached 0.979 area under the receiver operating characteristic curve and recovered 89.4% of seizure windows at a fixed 5% false-positive rate, with every one of the six unseen children exceeding 0.97.

• Pooled across all held-out windows the detector reached 0.979 area under the receiver operating characteristic curve and 0.650 area under the precision-recall curve against a seizure prevalence of 0.027, a 23.7-fold lift over chance. Macro-averaged across the six children the values were 0.987 and 0.816.
• At an operating point fixed to a 5% false-positive rate, sensitivity was 89.4% pooled and 93.9% macro-averaged, ranging from 83.9% to 100% across individual withheld children.
• Two cheap time-domain measures carry almost all the signal: signal variance contributed a summed importance of 0.598 and line length 0.259, together roughly 86% of the total. Among spectral features, relative gamma (0.028) and beta (0.026) power led.
• Predictive information concentrates on specific derivations. The two midline channels scored 0.100 and 0.092 and a right-temporal derivation 0.086, while prefrontal channels contributed least at about 0.029 each, consistent with their vulnerability to eye and muscle artefact.
• Discrimination is uniformly high across unseen children (0.973 to 0.998) but precision-recall performance varies substantially, from 0.605 to 0.977 — and that spread does not simply track prevalence. The child with the lowest seizure prevalence at 0.73% still reached 0.905, a 124-fold lift, with 100% sensitivity.
• Feature rankings were stable regardless of which child was withheld, indicating the model relies on generalisable electrographic correlates rather than subject-specific signatures.

How it was done

Six children from the Children's Hospital Boston and MIT database were analysed, sharing a consistent 18-channel bipolar montage at 256 hertz and contributing 30 clinician-annotated seizures totalling 47.0 minutes across 48 recordings. Signals were notch-filtered at 60 hertz, band-pass filtered from 0.5 to 45 hertz, and segmented into 4-second windows with 50% overlap. Ten features per channel gave 180 per window: variance, line length, skewness, kurtosis, five relative band powers, and spectral entropy. To prevent contamination, only windows lying entirely inside an annotated seizure were kept as positives, boundary windows were discarded, and positives and negatives came from disjoint files, with an audit confirming zero violations. A 400-tree random forest was evaluated under six-fold leave-subjects-out cross-validation. The 2.75% class imbalance was handled by downsampling the majority class to a 10-to-1 ratio and applying balanced class weighting on training folds only, leaving each held-out child's natural prevalence intact, with standardization fit on training folds alone.

Data sources

• CHB-MIT Scalp EEG Database version 1.0.0 on PhysioNet — 6 children, 48 recordings, 49,959 labelled windows of which 1,372 are seizure
• Shoeb & Guttag, Proceedings of ICML (2010) — the foundational patient-specific detection work on this database
• Beniczky et al., Clinical Neurophysiology 132:1173-1184 (2021) — clinical practice guideline on automated seizure detection with wearables
• Esteller et al., IEEE EMBS Proceedings (2001) — line length as an efficient seizure-onset feature