Problem
Researchers who test anti-epileptic drugs record the brain signal (ECoG, electrodes on the cortex) of epileptic rats for 6 hours after each dose. An expert then marks the start and end of every epilepsy episode and every deep-sleep episode by hand and sums the seconds to compare with the pre-drug baseline. Manual marking of 6-hour records takes a lot of expert time, so the customer wanted fully automatic labeling.
Example from the article: a model scored F1 0.67–0.93 on three test rats and was handed over, then failed on rats from other families. All three test rats were from one family.
Solution in brief
Cut each recording into 1-second windows, describe each window by wavelet-spectrum statistics, and feed a window together with its neighbors to a gradient boosting model, one model per episode type. Validate by holding out whole rats from different families, then remove the features that identify an individual and tune wavelet smoothing and context length per episode type.
Questions this case answers
- How to automate labeling of episodes in long multichannel signal recordings that experts currently mark by hand
- How to validate a window-based classifier so that its reported quality is not inflated by subject-specific patterns
- How to automate expert labeling of episodes in long biomedical time series
- How to validate a signal classifier so that it does not learn individual subjects
Tools and methods
- Wavelet transform — time-frequency features; Fourier does not suit non-periodic signals
- Context windows — features of neighboring seconds concatenated, prediction for the central second only
- Gradient boosting (CatBoost) — one model per episode type; beat transformer and CNN
- Leave-one-rat-out cross-validation — one rat is the test set, the rest train
- Label post-processing — drops very short labels, merges close ones
Google or a one-time email code. No password.