Machine learning · Research
EEG Data Analysis
Agglomerative hierarchical clustering, implemented from scratch on the Bangalore EEG Epilepsy Dataset, with a decision-tree surrogate that explains every cluster.
- Year
- 2025–26
- Role
- Research & implementation
- Built with
- Python, NumPy, Clustering, Decision trees
- Source
- Repository
the question
Clustering EEG recordings can surface structure in the signal, but a cluster label on its own explains nothing. On the Bangalore EEG Epilepsy Dataset the useful question is not only which recordings group together, but why.
the method
- 01
Worked with the Bangalore EEG Epilepsy Dataset, preparing the recordings for unsupervised analysis in Python and NumPy.
- 02
Implemented agglomerative hierarchical clustering from scratch rather than calling a library, so every merge step is visible and understood.
- 03
Built the cluster hierarchy bottom-up and assigned cluster labels from it.
- 04
Trained a decision-tree surrogate on those labels to explain the cluster assignments as readable rules.
- 05
Measured how closely the surrogate agrees with the clustering, and reported it honestly as label agreement, not diagnosis.
what it found
A from-scratch clustering pipeline with an interpretable layer on top. The surrogate reports 96–97% accuracy: agreement with the project's own cluster labels, not clinical diagnostic accuracy.
96–97%
surrogate agreement
0
EEG dataset, end to end
0
models: clusterer + surrogate
signal chain