Rationale: There is a shortage of neurologists in high and low-income countries.1 We, therefore, sought to develop an easy to use decision-making tool to help non-specialists diagnose epilepsy. Methods: Two physicians completed retrospective chart review of 214 consecutive patients aged ≥18 years with routine EEG data from MGH between 2016-2019. Exclusion criteria included EEG from a MICU admission, ultimate diagnosis of toxic-metabolic encephalopathy, and inability to determine the reason for EEG or the clinician’s diagnosis. We tabulated demographic information, reason for EEG, semiologic features of the presentation, final diagnosis, and EEG features. Patients diagnosed with epilepsy were compared to other diagnoses using t-tests for means, and chi2tests for frequencies. A multivariable logistic regression model was trained to predict the diagnosis (Epilepsy, Undetermined, or Non-epilepsy) using clinical variables and EEG information. We used ten-fold nested cross validation with LASSO for feature selection. Missing values were imputed with group medians. We fit a model with clinical predictors only, then simplified it by combining related predictors and eliminating predictors with small coefficients. To evaluate the diagnostic value of EEG findings, we trained a final model with EEG findings added. Results: Among 214 patients included, 112 were female (52.3%), with mean age 46.2 years (SD = 20.4). Diagnoses were epilepsy (45.3%), undetermined (28%), syncope (7.5%), migraine disorder (5.1%), other (5.1%), single unprovoked seizure (3.7%), psychiatric disorder (1.9%), stroke or TIA (1.4%), or provoked seizure (0.9%). Epilepsy patients compared to other differed by age (41.7 vs 49.8 years, p = 0.004); prevalence of developmental delay (15.6% vs 0.8%, p < 0.001), post-ictal confusion (43.8% vs 21.2% y, p < 0.001), witnessed convulsion (64.6% vs 22%, p < 0.001), abnormal ECG (5.2% vs 8.5%, p = 0.006), epileptiform discharges (24% vs 2.4%,), generalized spike and wave discharges (13% vs 5.6%, p < 0.001), and focal slowing (34.4% vs 25.2%, p < 0.001). Modeling identified 11 clinical features (Table 1), subsequently reduced to 5 features with these points: Substance abuse (-1 point), Stress induced (-1 point), GTC or Forced head turn (3 points), Syncope signs (-2 points), and Neuro injury (2 points); scores range from -3 to 5. The model area under the ROC curve was 0.84 [CI 0.78, 0.89]; calibration error was 0.06 [CI 0.02, 0.10] (Figures 1A, 1B). The model assigns “Undetermined" cases (n=12) intermediate probabilities (Figure 1C). We computed pre- and post-EEG probabilities for scores from -3 to 5 (Figure 1D). Absence of EEG findings decreases the post-test probability, while a positive result increases the post-test probability much more, with greatest impact when pre-test probability is near 40%. Conclusions: Our work suggests that a decision-making tool using a small number of easily obtained historical clinical features can provide accurate pre- and post-EEG probabilities of epilepsy. Non-specialists may be able to easily apply this tool in communities with limited access to neurologic care.
1 Burton A. How do we fix the shortage of neurologists? Lancet Neurol. 2018;17(6):502-503. doi:10.1016/S1474-4422(18)30143-1 Funding: Please list any funding that was received in support of this abstract.: NIH Click here to view image/table