Abstracts

Rationale: Controlled prospective clinical trials are generally agreed to be the most rigorous method for clinical research; however, they also may have weaknesses such as enrollment of very selective populations, narrow research focus, and long trial run time, which are most obvious in studying rare disorders with small patient populations. Practical alternatives may be observational studies using medical and administrative databases routinely generated in healthcare systems. However, the quality of administrative database studies is affected by several factors including coding algorithms for identifying patients with the disease of interest. This study aimed to validate the accuracy of algorithms to detect epilepsy cases using health records of a tertiary children’s hospital. We also sought to select an optimal strategy for identifying epilepsy cases. Methods: We performed a retrospective observational study of patients aged 0-19 years using clinical and claims data from the Children’s Healthcare of Atlanta (CHOA) from 1/1/2014 to 12/31/2014. Epilepsy case identification algorithms were classified into nine categories by combining the type and number of ICD-9-CM diagnosis codes for epilepsy and seizures (i.e., 1 code 345.xx; =2 codes 345.xx; 1 code 345.xx and =1 code 780.39; or =2 codes 780.39 and no 345.xx), encounter dates intervals between the diagnosis code detection (i.e., =1 day and < 30 days; or =30 days), and anti-epileptic drug (AED) prescriptions (i.e., yes or no). We defined putative non-epilepsy cases as follows: patients with 1 code 345.xx and no AEDs, or patients who were randomly selected from the rest except those included in the above-mentioned case definitions. Accordingly, the CHOA outcome center team extracted the putative epilepsy and non-epilepsy cases from claims and encounters data. To evaluate the validity of the epilepsy case definition, an experienced epileptologist thoroughly reviewed all patients’ electronic medical records (EMR) including clinic and inpatient notes of neurology and other providers, brain imaging results of CT and MRI, EEG results from routine, ambulatory, and long-term EEG monitoring, and medication lists for AEDs. The validity of case identification algorithm was assessed using sensitivity, specificity, positive predictive value (PPV), sensitivity and false positive rate (FPR). Results: A total of 1,802 putative epilepsy (n = 1,141) and non-epilepsy (n = 789) cases were detected by the algorithms. The EMR review revealed 1,145 epilepsy cases and 657 non-epilepsy patients. Positive predictive values (PPVs) ranged from 67.7% to 87.1%. Algorithms without AEDs had lower sensitivity and PPVs. The group of algorithms with both diagnosis codes 345.xx and AEDs had the highest PPVs (87.1%) with a sensitivity of 65.0% and an FPR of 16.7%. Conclusions: In this population, the PPVs were higher in algorithms defined by ICD-9-CM code 345.xx and AEDs prescription. Considering the low sensitivity and PPVs in algorithms without AEDs, it is recommended that both epilepsy diagnoses and AED prescriptions be used to identify epilepsy cases.   Funding: Children’s Healthcare of Atlanta research grant funded by the Goizueta Foundation.