The repo is an exploration of ML methods to classify EEG signals with siezures present/absent.
Dataset - CHB-MIT Scalp EEG Database
- 23 pediatric epilespy patients (1 subject has two datasets)
- Contious EEG recordings 1-4 hours long
- 198 labeled seizures
- 256 hz @ 16 bit resolution
- location: https://physionet.org/content/chbmit/1.0.0/
This is the result of a Random Forest model trained on a balanced data set, where siezures and non seizure epochs equally represented. Various statstics (peak-to-peak,variance,mean,zero-crossings) from each channel served as features. Just a quick sanity test to get up and running.
Figure 1. Histogram of raw data from each class. Qualitative differences suggest a model should be able to classify.
Figure 2. Confusion matrix. The model performs quite well on the balanced data.