EEG analysis for harmful brain activity

I built a framework to classify EEG patterns associated with harmful brain activity, working from raw signals and spectrogram representations. The goal was practical detection under real clinical noise, not just leaderboard accuracy on a clean slice of data.

I compared several CNN architectures (EfficientNetB2, MobileNetV3, ResNet V2, DenseNet) in TensorFlow/Keras, with preprocessing that included normalization, log transforms, and standardization. EfficientNetB2 reached 81.92% accuracy. I also used Kullback-Leibler divergence for probability modeling and built retrieval pipelines with PySpark and Parquet to speed up batch analysis.

Focus

• EEG and spectrogram classification at scale
• CNN architecture comparison for clinical signal data
• PySpark pipelines for faster batch inference