Abstracts

Rationale: Distinguishing epileptic (ES) from non-epileptic seizures (NES) in an outpatient setting is challenging due to their seizure behavioral similarities. Improved methods to reach the diagnosis objectively without inpatient video-EEG monitoring would reduce the time to diagnosis, and initiation of treatment; this also may decrease the cost of evaluation. We examined if patients' patterns of comorbidities and pharmacologic treatment history, derived solely from archived clinical notes, could help distinguish ES from NES. Methods: To quantify the diagnostic utility of comorbidities and pharmacologic treatment history, we studied epileptologists' initial outpatient notes for 298 patients with medication resistant seizures (218 ES, 80 NES, 0 NES+ES). These patients subsequently were diagnosed with either ES or NES based on inpatient video-EEG monitoring. Not all factors were reported in all patients, leading to some missing data. Patterns of comorbidities and pharmacologic treatment associated with ES vs. NES were determined using an entirely data-driven C4.5 decision tree using cyclical leave-one-out cross-validation. All negative results on our analysis indicate NES. Results: Based on the patient encounters, our C4.5 decision tree achieved 79% leave-one-out cross-validation accuracy (95% confidence interval 74-84%: better than chance, permutation test, p<0.001). While our method had high sensitivity (90%), its specificity was lower (49%). The positive and negative predictive values were 83% and 64%, respectively. The learned C4.5 tree was 12 nodes deep, and did not include gender as a risk factor despite the finding that 74% of patients with NES were female. Instead, the data-driven tree prioritized the number of seizure and non-seizure medications, as well as the presence and/or treatment of psychiatric comorbidities. Conclusions: This good accuracy, and excellent sensitivity, suggests that our decision tree could be used to identify ES reliably. The low specificity and negative predictive value, however, suggests that its ability to identify NES was limited. Therefore, our algorithm could be applied most effectively to medication-resistant patients as a method for ruling out the possibility of NES. In addition to the overall results, the structure of the learned decision tree provided insight that could be interpreted to better illuminate the medical and psychiatric causes, effects and associations of seizures. For instance, our tree suggests that the gender difference in patients with NES can be accounted for by the presence of psychiatric comorbidities. This is consistent with the mechanistic theory that NES are signs of conversion disorder; female gender itself does not increase the risk of NES. This work could be used to identify better which patients may benefit from anti-seizure medications, as well as identify key non-seizure comorbidities that, if left untreated, could reduce patients' quality of life even if their seizures are controlled. Funding: NIH (T32 GM08042, T32 GM008185, T90 BA023422, R90DA023422), William M. Keck Foundation, UCLA Dept. Biomath