Hosted by Harvard Medical School | Research Code Competition | Silver Medal (Rank: 87th out of 2767 teams, top 3.1%)
This competition seeks to detect and classify epileptic seizures and other pathological brain activity through automated analysis of electroencephalogram (EEG) data. The project holds particular significance for enhancing the efficiency and accuracy of EEG monitoring in intensive care unit (ICU) settings. The underlying problem constitutes a standard classification task, and the competition employs the Kullback-Leibler divergence as its primary evaluation metric.
Initially, we extracted 17,089 training samples from the competition dataset, using unique EEG identifiers as the selection criterion. Based on these samples, we identified the following features for model development:
1）Spectrograms provided by the competition organizers.
2）Mel-spectrogram features extracted from the raw EEG data using the librosa package.
By concatenating these two feature sets, we generated composite spectrogram images for subsequent analysis.
Drawing from experience in image processing, we considered three common computer vision model architectures: ResNet34, EfficientNetB0, and EfficientNetB1. These models have demonstrated outstanding performance in various tasks and are well-suited to the characteristics of the competition data. To enhance the robustness of the models, we employed group k-fold cross-validation, splitting each patient's samples into five folds to thoroughly test and mitigate overfitting.
Finally, we won a silver medal and ranked 87th out of 2767 teams.
