A 2.7 MB AI detects childhood seizures, and its biggest contribution is explaining why it fails in some children

By Bioengineer.org · October 16, 2026.

A New Zealand team led by Raha Pedram and Hamid Abbasi has published a deep learning model in Results in Engineering that detects epileptic seizures in children from raw electroencephalograms (EEG). According to the authors, it is a temporal convolutional network (TCN) with 709,697 parameters, about 2.71 MB.

The model was trained on CHB-MIT, a dataset from Boston Children's Hospital and MIT covering 23 pediatric patients across 24 recording sessions. It processes 18 EEG channels directly, with no band filtering, normalization or denoising. Each training epoch takes about three seconds on an NVIDIA A100 GPU.

The headline figure is deliberately modest. Under strict validation, with one full session held out for testing in each of the 24 rounds, the model achieved a mean accuracy of 82.3% (standard deviation 16.5), a sensitivity of 80.1% and an AUC of 0.866. A Transformer model (Wav2Vec2), used as an internal benchmark, reached 76.9% under the same conditions.

Under a different validation scheme, which keeps each case's windows grouped together, accuracy rises to 98.2% and AUC to 0.999. The gap between the two figures is a finding in itself. The authors argue that many earlier studies on CHB-MIT excluded difficult cases (12 to 16) or over-filtered the data, which inflates results.

By case, 15 sessions performed well (above 80% accuracy), 5 moderately and 4 poorly (below 65%). In two of them, chb13 and chb14, the AUC fell below chance. Using only the 15 good cases, accuracy was 91.5%, although the authors caution that this analysis is exploratory and does not amount to independent validation.

Spectral analysis provides the explanation. Seizures stood out mainly through simultaneous power increases in the delta (0.5-4 Hz) and theta (4-8 Hz) bands. In well-classified cases, delta rose more than fivefold and theta three- to fourfold relative to baseline. The theta difference showed the strongest correlation with per-case performance (Spearman's rho of 0.578; adjusted p of 0.0083).

In the difficult cases, the opposite occurred. Some had elevated baseline delta outside seizures, which produced false positives. In others (chb12 to chb15), seizures showed no clear low-frequency increase, resulting in false negatives. High-amplitude noise was also observed in certain recordings. The authors warn that the dataset lacks ocular channels and sleep or movement annotations, so these patterns are signal observations rather than proven physiological markers.

The most relevant finding for real-world use comes from the evaluation on continuous recordings with the natural proportion of seizures. The model detected 91.0% of events, with a mean latency of 15.9 seconds. But it generated 3.21 false alarms per hour and an alarm-level positive predictive value of just 3.4%. The authors themselves state that it is not suitable for clinical deployment until false alarms are reduced.

**Our take.** Our view is that the useful advance here is not the score but the methodological honesty. An 82% under per-session validation counts for more than a 99% with mixed windows. As industry context, this is a rule worth applying to all medical AI: ask what population and what data split it was validated on.

The second lesson is that inter-patient heterogeneity matters more than architecture. If it is possible to predict, from their delta-theta signature, which children will be hard to read, the logical next step is case-adapted models. That remains to be seen, because with 23 patients from a single dataset, generalization remains unproven.

An alarm that is right 3.4% of the time will wear out any family or nursing team, and that is the real short-term obstacle. In the long run, however, a 2.7 MB detector fits on a low-cost wearable device. That points to accessible diagnosis and monitoring in low- and middle-income countries too, where, according to the article, epilepsy incidence is higher. It is a small but well-directed step toward more universal medicine.