Safe Seizure - Brain seizure prediction
Safe Seizure: Brain seizure prediction using machine learning
Language: Python
1 Problem
Around 50 million people on Earth have chronic epilepsy. In about 40% of cases, patients have drug-resistant epilepsy against which anti-seizure treatments have no effect.
Being able to predict, well in advance, when the next seizure will occur could significantly improve epileptic patients lives.
2 Solution: Safe Seizure
Using Machine Learning, the Safe Seizure team is able to predict if a seizure will occur up to 3 hours in advance using intracranial EEGs (iEEGs).
3 Objectives
-
Objective 1: Distinguish preictal (i.e. pre-seizure) from interictal (i.e. no-seizure) states –> binary classification
-
Objective 2: For preictal (i.e. pre-seizure) states, predict if a seizure will occur in 1, 2, or 3 hours –> multi-class classification
4 Results
| Results 1: overall accuracy of 84% | Results 2: overall accuracy of 87% |
|---|---|
 |
 |
In Results 2, the mis-classifications that we absolutely want to avoid are circled in red.
Also, in orange rectangles show a confusion between 2 and 3 hours prior to seizure.
5 Algorithm
The Machine Learning algorithm developed for the project is inspired by al-Qerem et al. 2020.
Below is an outline of the sequential algorithm:
- The raw iEEG time series are denoised using Independent Component Analysis (ICA).
- The denoised signals are then filtered using a Discrete Wavelet Transform (DWT).
- Statistical indicators are extracted from the filtered signals (mean, std, entropy, mav…)
- Classification using a Random Forest Classifier (RFC)
