Research Proposal

Using a Long-Short Term Memory Network to Detect Seizures in Epileptic Patients

Lay Description

For epileptic patients in particular, seizures can be dangerous and unpredictable. Electrical activity from the brain is often used for detecting seizures because of the drastic electrical spikes that occur before a seizure. However, these can be difficult and time-consuming to use on a regular basis. My project takes a different approach by employing artificial intelligence (AI) and heart signals (ECG) to more readily and precisely identify seizures. I created an LSTM (Long Short-Term Memory) network, a specialized AI model, that can identify seizure patterns by examining how the electrical activity of the heart of an epileptic individual changes minutes before a seizure. By balancing the various data kinds, standardizing the signals, and producing meaningful sequences for the AI to learn from and enhance its performance in the future, this model was able to learn from actual patient data. To make the model more reliable, I use techniques like dropout, regularization, and focal loss, which help prevent errors and improve accuracy. I also plan to fine-tune the AI by testing different settings (hyperparameter tuning) and enhancing its ability to explain its decisions using SHAP (SHapley Additive Explanations), a method that helps doctors understand why the AI made a certain prediction. The overall goal of my project is to create a fast, accurate, and explainable AI system that can help doctors and patients detect seizures using ECG data, potentially leading to wearable seizure monitoring devices that will improve patient safety and quality of life.

Grant Proposal/Research Plan

Project Notes

Pictures of Work - Preliminary Data

Accuracy Over Epochs for an LSTM networkAccuracy Over Epochs for a CNN model

For stability and controlled learning, the LSTM seems better because its accuracy and loss curves indicate smooth learning without erratic fluctuations.
For the CNN, both accuracies reach near 100%, which may suggest overfitting. This is unusual for real-world data and suggests memorization rather than generalization.
The early epoch instability in validation accuracy, with sharp fluctuations, indicates that the model struggles to generalize and may be adapting too much to the training data.