Abstracts

Rationale: Lennox–Gastaut syndrome (LGS) is an epilepsy disorder often under-diagnosed partly due to its heterogeneous clinical profile, comprised of a myriad of symptoms. Health insurance claims databases offer a wealth of data to understand diseases with low prevalence, but currently lack specific diagnosis codes for LGS. Valid approaches for identifying LGS in claims databases are needed. This study aimed to develop a claims-based classifier to identify patients (pts) with LGS using a random forest (RF) methodology.Methods: Clinical experts in LGS were consulted to identify potential predictor characteristics of LGS in a claims database. Pts with epilepsy (≥2 claims for ICD-9 345.xx) from 6 state Medicaid programs (1997–2013) were assessed and categorized into LGS cases, defined as pts with ≥1 filled prescription for rufinamide (indicated solely for LGS), vs non-LGS cases, defined as pts with only non-intractable epilepsy (ICD-9 345.x0). LGS and non-LGS cases were randomly split into 2 datasets (training vs testing) to develop and validate a RF classifier for LGS. The RF classifier was constructed in the training dataset. Sensitivity and specificity were then computed in the testing dataset to assess the performance of the classifier in identifying LGS cases. Multivariate logistic regressions using the top 15 most important predictor characteristics as ranked by the RF classifier were also conducted to provide a conventional context to assess the statistical significance of these characteristics.Results: Compared to other epilepsy pts (N= 258,503), rufinamide-treated cases (N= 1,206) were younger at 1st observed epilepsy diagnosis (mean [SD]: 9.2 [10.3] vs 32.1 [22.6] years), used more antiepileptic drugs (7.4 [2.8] vs 2.2 [2.0]), and had more all-cause (7.2 [9.9] vs 4.1 [7.4]) and epilepsy-related (3.0 [5.3] vs 0.9 [2.2]) outpatient visits annually. Mental retardation/delayed development (91.5% vs 33.8%), vagus-nerve stimulation implant (33.9% vs 1.7%), West syndrome (14.7% vs 0.9%), and helmet use (14.3% vs 0.7%) were more common among rufinamide-treated pts. The best-performing RF classifier yielded a sensitivity of 97.3% and a specificity of 95.6%. The top 15 most important predictor characteristics of LGS status are listed in Table 1. Logistic regression results indicated that most of the top predictor characteristics were significantly associated with LGS status (p-values <0.05); exceptions were epilepsy-related inpatient visits, number of neurological procedures, use of wheelchair/walker, and West syndrome.Conclusions: The current claims-based LGS classifier is comprised of clinically and statistically relevant predictor characteristics, and yields high sensitivity and specificity. Identification of LGS pts in claims data will enable the study of this under-recognized disorder in real-world settings. Funded by Lundbeck, LLC