A comparison of statistical versus clinical classifiers of seizure clustering in women with catamenial and noncatamenial epilepsy

Abstract

AbstractObjectiveWe assessed whether (1) women with statistical clustering of daily seizure counts (DSCs) or seizure intervals (SIs) also showed clinical clustering, defined separately by ≥2 (≥2‐SC) and ≥3 (≥3‐SC) seizures on any single day; and (2) how these classifiers might apply to catamenial epilepsy.MethodsThis is a retrospective case–control analysis of data from 50 women with epilepsy (WWE). We assessed the relationships of the four classifiers to each other and to catamenial versus noncatamenial epilepsy using chi‐squared, correlation, logistic regression, and receiver operating characteristic (ROC) analyses.Results≥3‐SC, not ≥2‐SC, was more frequent in WWE who had statistical DSC clustering versus those who did not (21/25 [84.0%] vs. 11/25 [44.0%], p = .007). Logistic regression (p = .006) and ROC (p = .015) identified ≥3‐SC, not ≥2‐SC, as a predictor of statistical DSC clustering, but ≥4‐SC was more accurate. ≥3‐SC correlated with the average daily seizure frequencies (ADSFs) of the subjects (p = .01). ROC optimal sensitivity–specificity cut‐point for ADSF prediction of ≥3‐SC (.372) was 64.6% higher than for ≥2‐SC (.226). SI clustering was more common in WWE who had catamenial versus noncatamenial epilepsy (p = .013). Logistic regression identified statistical SI clustering as the only significant classifier (p = .043). ROC analysis offered only marginal support (p = .056), because specificity was low (42.1%).SignificanceThe findings lend statistical support for (1) the utility of clinical ≥3‐SC as a predictor of convulsive status epilepticus, (2) consideration of ADSFs in defining clustering, and (3) ≥4‐SC as a more accurate clinical predictor of statistical DSC clustering. Statistical SI clustering occurred more frequently in women with catamenial than noncatamenial epilepsy (90.3% vs. 57.9%, p = .013). Although sensitivity was high (90.3%, 28/31), specificity was only 42.1% (8/19). Algorithms that test patterns and periodicities of clusters are more applicable.