ESL RF AZCfeat 2023
- Developers: Renato Zanetti, Una Pale, Tomas Teijeiro, David Atienza
- Institution: ESL lab, EPFL, Switzerland
- Contact: una.pale [at] epfl.ch
- Source: https://github.com/esl-epfl/AZCfeature
- License: MIT
- Reference paper: https://iopscience.iop.org/article/10.1088/1741-2552/aca1e4
- citation: DOI 10.1088/1741-2552/aca1e4
Model details
Signal Processing pipeline:
Features are extracted from raw data without preprocessing. We postprocess predicted labels using Bayes approach that calculates accumulative probability of a seizure based on probability of individual predictions. We also have tolerance for resolution of prediction around true annotations.
Model Description
Random forest model with 100 trees (maximal three depth 0 and Gini impurity for splitting criteria).
Model Features
We use 4s long windows which are shifter in steps of 0.5s. We used approximate zero crossing features which are based on polynomial approximation of the signal with different thresholds and then counting number of zero crossings. In total we have 6 AZC features.
Model Channels
19 channels
Model Training
We used at least 5h of data to start to train, or at least one seizure. Then we tested using time-series-cross-validation where we always tested on the next 1h of consequent data. And for every new CV we enlarged training data with 1h of test data from previous step of CV. In the end we appended all test sets.
Model Imbalance
We didn’t deal with it. We used all data as it is.