Machine Learning Analysis of Spectro-Graphic Data for Biomarker Identification

Abstract

Epilepsy diagnosis often relies on manual interpretation of EEG data, a process that is time-consuming, prone to error, and difficult to scale. Our project addresses this by developing a machine learning model capable of detecting epileptic events in EEG signals with accuracy and speed. We trained the model using a combination of internally collected EEG data (96 samples) and publicly available Bonn dataset segments (450 samples). After optimizing hyperparameters, the model was validated on unseen test data and achieved an overall accuracy of 97.03%. It demonstrated perfect precision (1.0) and 98% recall in detecting seizures, and 98.6% precision with 93% recall for pre-seizure activity, indicating strong predictive power for early warning. Normal EEG segments were identified with 100% recall and 92.2% precision. To simulate clinical use, we developed a real-time processing script that categorized incoming EEG samples within 10 seconds, mimicking continuous monitoring workflows. This validates the model’s feasibility for real-world applications, including portable or at-home diagnostic tools. With the AI healthcare market projected to grow at a CAGR of 45% from 2023 to 2033, and a global rise in diagnosed neurological disorders, our solution directly addresses a critical clinical and commercial need. Future work will focus on refining the model’s performance across diverse populations and expanding real-time capabilities.

Video