Despite the widespread adoption of new drug classes to treat epilepsy, approximately one third of patients do not respond to anti-epileptic medication. Patients with drug-resistant epilepsy (DRE) suffer permanent memory and cognitive impairment from uncontrolled seizures. Only 8-10% of patients with DRE obtain long-term seizure freedom. About half of patients with DRE are eligible for surgical treatment, which results in seizure freedom in 58-73% of patients. However, it is difficult to determine when patients with DRE need surgery. DRE has a complex disease course that fluctuates according to weekly and monthly patterns. Further, neurologists report that they lack the resources (i.e. time) needed to analyze and interpret large amounts of electronic health record (EHR) data to determine surgical candidacy. This results in substantial delays in treatment - 6 years in pediatrics and 20 years in adults, on average - and contributes to avoidable morbidity and mortality. Automated systems are capable of assisting clinicians in identifying candidates for epilepsy surgery two years earlier in the disease course. One such system uses natural language processing to analyze free-text neurology notes in a real-time clinical setting. This system was able to increase the rate of surgical candidate identification by 46%, but it could still be improved in two important ways. First, the system is not able to incorporate results from EEG and MRIs - the most influential factors on surgical candidacy - into its recommendations. Second, it does not utilize rich information hidden in structured EHR data that captures epilepsy disease burden. A system that fuses information from all three data sources (neurology notes, EEG and MRI reports, and structured data) would drastically improve the accuracy and impact of the model. In the proposed research, I will first develop a deep learning (DL) approach for fusing multi-modal EHR data. Specifically, I will use recent advances in deep representation learning to produce richer features from both free- text and structured data. I will represent free-text in neurology notes and EEG and MRI reports with word embedding vectors, and medication, procedure, and visit codes with medical concept vectors. I will combine disparate data sources with a deep neural network to produce high-level representations of surgical candidacy. This will enable me to estimate patients' risk of future epilepsy surgery. Second, I will establish the generalizability of this approach using a neighboring hospital's EHR data and validate the system in a clinical setting. This contribution will be significant because it will increase the number of surgical candidates identified earlier in the disease course, thereby reducing epilepsy disease sequelae. This proposal will lay the groundwork for nationwide expansion of the DL system and generate the only automated DL system designed to improve the timeliness of surgical referral rates in the adult population.
Automating the identification of patients with epilepsy who would benefit from surgical treatment is imperative to reducing morbidity and mortality from uncontrolled seizures. This research proposal will develop and clinically validate a generalizable deep learning framework for analyzing multi-modal electronic health record data to predict epilepsy surgery candidacy. Implementing this system into clinical care is expected to reduce the time to surgery by 45%, which would prevent one epilepsy-related death for every 58 patients who receive surgery.
