Sonali Santhosh
Class of 2026Richland, WA
About
Hello! My name is Sonali Santhosh and I decided to work on this project because I’m deeply fascinated by the brain, especially seizures, and how they affect our most vulnerable patients. As someone who aspires to become a neurosurgeon in the future, I’m particularly interested in understanding and addressing critical neurological conditions such as seizures. This project combines my passion for neuroscience with my goal of making an impact on patient care, and helped grow my understanding of machine learning.Projects
- "Cracking the Seizure Code: Leveraging Bi-LSTM Models for Neonatal EEG Interpretation and Seizure Classification" with mentor David (Oct. 5, 2024)
Sonali's Symposium Presentation
Project Portfolio
Cracking the Seizure Code: Leveraging Bi-LSTM Models for Neonatal EEG Interpretation and Seizure Classification
Started Mar. 29, 2024
Abstract or project description
This study explores the effectiveness of Bidirectional Long Short-Term Memory (Bi-LSTM) in predicting the onset of seizures in neonates using Electroencephalogram (EEG) data, based on expert neurologist diagnosis. The model was able to function as a universal classifier across multiple patients, aiming not to predict seizures directly but to closely correspond to the expert diagnoses reflected in the provided annotations. The model's key objective is to approximate human expertise, using these annotations as the target values rather than attempting to identify seizure events independently.
In an infant’s critical first moments, their brain is a whirlwind of activity, making it difficult to distinguish abnormal patterns like seizures, which can strike without warning and have devastating effects. Given the critical need for timely seizure detection in neonatal intensive care units (NICUs), traditional methods, which often rely on professional analysis, can be slow and less effective in early detection. To address these challenges in neonatal seizure detection, Machine Learning offers promising solutions. This study involved a publicly available dataset with 79 term neonates, the 79 EDF files were then processed by bad signal removal, band-pass filtering, segmentation in fixed-length epochs, three feature selections, and standardization. The Bi-LSTM model was trained throughout multiple epochs, followed by performance evaluation. The model was able to universally classify the individuals and obtain an overall accuracy of 80%. Out of all 79 neonates, the model could accurately predict 91% of the non-seizure activity, and 47% of the seizure activity. A key limitation present in this study is that of no concrete “ground truth” as even the expert annotations are subjective. This caused the model to have no guarantee that it was trained and tested on true information, leading to the wide range of accuracy present between individuals. Furthermore, overlap between the visual characteristics of seizure and normal brain activity in certain individuals may also cause a lack of accuracy in model seizure prediction.
Thus, this study determines the potential of Bi-LSTMs in accurately detecting neonatal seizures, based on EEG data, showing promise for reducing the long-term impacts and improving outcomes for vulnerable neonates by aligning closely with expert annotations.