Associate Professor Cincinnati Children's Hospital Medical Center
This abstract will be presented during the Clinical Epilepsy/Antiseizure Medication Platform poster session
Rationale: Epilepsy affects more than 479,000 children. The average interval from diagnosis to epilepsy surgery in pediatrics is six to ten years. Early identification and referral of children who are potential surgical candidates is a laborious and complex process. Natural Language Processing (NLP) and machine learning techniques have been successfully used to evaluate clinical notes and make recommendations; however, they frequently are not implemented into care. The objective of this study was to implement and test an existing NLP-based automated clinical decision support system to identify patients who may be eligible for neurosurgical evaluation. Methods: The prospective, randomized controlled trial, took place at a large, pediatric epilepsy center (Cincinnati Children’s Hospital Medical Center; Cincinnati, Ohio, USA). All patients with upcoming visits in the neurology clinic were screened by an existing NLP algorithm1 before their scheduled appointment. Patients who were classified as a potential surgical candidate were randomized to have their provider receive an automated email, an alert in the electronic health record (EHR), or no alert (control group). Randomization occurred at the visit level. Patients were followed for one to three years after randomization. Patients who were not referred within six months of their randomization were eligible to undergo a second randomization at subsequent visits. The primary outcome was referral for a presurgical evaluation. Likelihood of referral was assessed using a Cox proportional hazards regression model. Results: Patients were enrolled from April 16, 2017 to April 15, 2019 and followed for one to three years. There were 93 automated emails, 95 EHR alerts, and the remaining 96 visits without reminders were standard of care (Figure 1). There were no significant differences in demographics. Compared to patients in the control group, patients who received an automated alert were more likely to be referred for a presurgical evaluation (3.1% vs. 9.8%; HR = 3.12 [95% CI: 0.93 - 10.5]; p = 0.04, likelihood ratio test) (Figure 2). Nine patients (4.4%) underwent neurosurgical treatment after receiving an automated alert, compared to none (0%) in the control group (p = 0.06). Conclusions: Clinicians were more likely to refer epilepsy surgery candidates for a presurgical evaluation after receiving an automated alert from an NLP-based clinical decision support system. The intervention can help providers identify patients who are eligible for neurosurgical evaluation earlier in the disease course. Funding: Please list any funding that was received in support of this abstract.: National Institutes of Health (K25HL125954) and the Agency for Healthcare Research and Quality (R21 HS024977).