Leveraging Clinical Data and Automation to Predict Epilepsy After TBI

About This Grant

PROJECT SUMMARY Post-traumatic epilepsy (PTE) affects nearly 400,000 patients in the US each year significantly impairing recovery and increasing psychological and financial burdens. Despite its profound impact, strategies for PTE prevention and prediction remain elusive, with current risk models hindered by their ability to scale to larger health systems or populations, reliance on data that is not readily available to the broad traumatic brain injured population or the need for labor-intensive extraction of important variables from the clinical heath record. These limitations restrict applicability of existing models to broader patient populations and exacerbate health disparities. This project aims to overcome these barriers by developing automated, scalable, and clinically applicable tools for PTE risk stratification. Our central hypothesis is that automated extraction and analysis of multidimensional clinical and imaging data can identify high-risk PTE patients, enabling early intervention and enhancing treatment feasibility. We will leverage advances in machine learning, natural language processing, and neuroimaging to analyze data from approximately 3,000 TBI survivors, addressing three specific aims: (1) Optimize Automated EHR Phenotyping: Develop an algorithm to retrospectively identify PTE patients to more easily identify a large cohort of PTE patient for future PTE risk prediction models and to use electronic health record (EHR) data from the first seven days post-TBI to predict future PTE risk. We will analyze structured and unstructured data to uncover novel predictors and word clusters indicative of PTE risk. (2) Quantify Contusion Features with Deep Learning: Implement a convolutional neural network (CNN) to automatically measure contusion volumes and locations from acute CT scans to assess their predictive value for PTE risk. (3) Evaluate Cortical Volume and Thickness from MRI: Use CNN-based segmentation of clinical MRI scans to analyze cortical volume and thickness to explore their association with PTE. As a final Exploratory Aim, we will integrate the features from these aims into a unified prediction model, hypothesizing that a comprehensive approach will outperform single-modality models. This work has the potential to transform PTE risk prediction by leveraging widely available clinical data and automated tools, providing a foundation for more effective diagnosis, prevention, and treatment strategies across diverse healthcare environments.