Feasibility of Using Regular Expressions to Evaluate SUDEP in Clinical Notes
Abstract number :
2.410
Submission category :
17. Public Health
Year :
2017
Submission ID :
345582
Source :
www.aesnet.org
Presentation date :
12/3/2017 3:07:12 PM
Published date :
Nov 20, 2017, 11:02 AM
Authors :
Kristen Berry, Weill Cornell Medical College, Cornell University, New York, NY, USA; on behalf of Pediatric Status Epilepticus Research Group (pSERG), Boston Children’s Hospital, Harvard University Medical School, Boston, MA, United States; Stephen
Rationale: Sudden unexpected death in epilepsy (SUDEP) is an important cause of mortality in epilepsy. The vast majority of families and adults with epilepsy want to be informed about SUDEP.1,2 However, a gap remains in how often SUDEP is discussed. One potential solution is to automatically prompt physicians in electronic medical records (EMRs) to counsel high risk patients. To do so requires automated detection of risk factors, which are not reliably captured in standard diagnostic codes. Regular expressions, a method of natural language processing (NLP), presents an opportunity to identify SUDEP risk factors by analyzing text of physician notes. The current study evaluated: 1) how SUDEP is discussed in clinical notes, and 2) the feasibility of using regular expressions to identify SUDEP risk cohorts. Methods: Data included physician notes in EMRs at an academic medical center from 2010-2014. We performed text searches to evaluate how SUDEP was documented in epilepsy clinic notes (n=4,716 patients) and inpatient neurology notes (n=8,957). Next, we evaluated the feasibility of using regular expressions, which identify simple patterns of characters in text, to identify risk cohorts. Among epilepsy clinic notes, we selected a training set (n=1,000 patients) and test set (n=1,000). We used manual coding and regular expressions to identify patients with: 1) possible/known seizures, 2) generalized convulsive seizures, 3) refractory epilepsy, and 4) epilepsy surgery candidates. We evaluated performance of regular expressions using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), compared to gold standard of manual coding. Results: SUDEP counseling was documented for 2 of 4,716 (0.04%) and 3 of 8,957 (0.03%) patients in epilepsy clinic and inpatient neurology notes, respectively. In our training set, 822 of 1,000 patients had possible or known seizures. Of those, 154 had convulsive seizures, 137 had refractory epilepsy, and 38 were epilepsy surgery candidates. Training set performance for possible/known seizures had 100.0% sensitivity, 98.9% specificity, 99.8% PPV, and 100.0% NPV. Performance for convulsive seizures had 93.3% sensitivity, 98.5% specificity, 92.6% PPV, and 98.6% NPV. Performance for refractory epilepsy had 96.2% sensitivity, 99.7% specificity, 98.1% PPV, and 99.3% NPV. Performance for epilepsy surgery had 92.9% sensitivity, 100.0% specificity, 100.0% PPV, and 99.7% NPV. Conclusions: Discussion of SUDEP is rarely documented in epilepsy clinic and inpatient neurology notes. Regular expressions are a feasible and potentially valuable tool to identify SUDEP risk cohorts in EMRs. Our ongoing work will (1) evaluate performance in a test set, and (2) expand analysis to additional institutions to evaluate broader SUDEP counseling and generalizability of NLP methods. References: 1. Epilepsia. 2010;51:777-782; 2. Epilepsy Behav. 2015;42:29-34. Funding: This work is supported through generous funding from (1) the Centers for Disease Control and Prevention (U01DP006089) and (2) Terwilliger Family Initiative for Innovation in Epilepsy Research Risk Detection and Prevention of Death Fund.
Public Health