MD-PhD Student Cincinnati Children's Hospital Medical Center Cincinnati, Ohio
Rationale: Resective epilepsy surgery has the potential to eliminate seizures in patients with drug-resistant epilepsy, but the average duration of disease prior to surgery is six years in pediatrics and 20 years in adults.1 Our objective was to develop a deep neural network (DNN) to identify candidates for epilepsy surgery earlier in the disease course. Methods: In this retrospective longitudinal cohort study, we used adult data from 2012-2019 and pediatric data from 2009-2019 from two large, academic epilepsy centers in Cincinnati, OH, USA. All patients with an International Classification of Diseases (ICD)-9 or -10 diagnosis of epilepsy and at least two neurology visits were eligible for inclusion. Cases were defined as patients who underwent resective epilepsy surgery, and controls were non-surgical patients. For surgical patients, only data prior to their presurgical evaluations were included. Free-text neurology notes and EEG and MRI reports were tokenized into one- to three-word strings (n-grams) and analyzed using support vector machines. Patient demographics, medications, procedures, labs, office visits, emergency department visits, and hospitalizations were analyzed using a random forest. Outputs from these models, plus the duration of neurology follow-up, were fused using an artificial neural network (Figure 1). The DNN’s output was the estimated likelihood of surgical candidacy. Model performance was measured using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (PR-AUC), which we obtained from ten-fold cross validation. Calibration of model estimates were analyzed using nonparametric calibration curves. Results: A total of 5,957 pediatric patients (51% male; 13.6 ± 7.48 years old) and 7,604 adult patients (56% female; 47.5 ± 16.8 years old) were included in the analysis. Pediatric and adult cohorts contributed a per-patient mean ± standard deviation of 6.4 ± 4.8 and 6.4 ± 5.0 observations, respectively. Corresponding durations of follow-up were 3.0 ± 2.6 and 3.0 ± 2.2 years. The DNN analyzed 180,602 and 151,373 variables extracted from the pediatric and adult EHRs, respectively. The surgical candidacy scores were consistent with the observed rate of surgery in both populations (Figure 2). AUROCs in the pediatric and adult datasets were 0.96 (95% CI = 0.95 to 0.98) and 0.93 (95% CI = 0.89 to 0.97), respectively, and PR-AUCs were 0.62 (95% CI = 0.58 to 0.66) and 0.49 (95% CI = 0.33 0.66). The DNN significantly outperformed a baseline logistic regression model that included age, race, gender, the number of anti-epileptic drugs prescribed, and the duration of follow up with neurology (PR-AUC was 40.6% higher in pediatrics and 42.3% higher in adults; p < 0.001 for both). Conclusions: The DNN generated calibrated surgical candidacy scores that had excellent discrimination in both pediatric and adult patients. These scores can be used to identify potential candidates for epilepsy surgery earlier in the disease course. Funding: Please list any funding that was received in support of this abstract.: Research reported in this abstract was supported by the National Institutes of Health (F31NS115447 and K25HL125954) and the Agency for Healthcare Research and Quality (R21 HS024977). Click here to view image/table