Abstracts

Rationale: The use of electronic medical record (EMR) data to build predictive models of disease using large datasets is becoming increasingly common. Stroke is one of the most common brain injuries and a substantial contributor to the prevalent epilepsy, especially in older adults. However, we lack means to automate EMR phenotyping of post-stroke epilepsy (PSE).

Methods: We sought to validate a method for identifying post-stroke seizures from EMR using the Cleveland Clinic Enterprise Data Vault (EDV), a data warehouse and medical data processing system updated daily with clinical and laboratory information from the EMR. In this study, we detected seizures in EDV encounter diagnosis data, defined as the presence of epilepsy ICD-9 (i.e. 345, 780.3) or ICD-10 codes (i.e. G40). We tested this method for the identification of acute symptomatic seizures (ASyS) and PSE. We used our cerebrovascular center’s prospectively maintained database of ischemic stroke admission 01/2014 - 10/2022. Patients with a pre-existing epilepsy ICD code in their encounter data prior to their stroke were excluded. The remaining patients were then labelled for presence / absence of ASyS and PSE in EDV based on the time to ICD code entry in the EMR (≤7 or >7 days, respectively). One hundred encounters in each group of interest (positive for ASyS, negative for ASyS, positive for PSE, negative for PSE) were then randomly selected and underwent chart review by a neurology resident to determine the ground truth. The EMR extraction method for ASyS and PSE was evaluated against the chart review using classification accuracy metrics, including positive predictive value (PPV), negative predictive value (NPV), overall accuracy, sensitivity, and specificity. Confidence intervals (95%) were calculated for sensitivity and specificity.

Results: Out of 6561 patients in the stroke database, 340 (5.2%) patients with epilepsy history were excluded. There were 274 unique patients with reviewed encounters; see Table 1 for demographic details. Using epilepsy ICD codes from EDV encounter tables led to ASyS PPV of 73.0%, NPV of 98.0%, combined accuracy of 85.5%, sensitivity of 97.3% [0.88 – 1.07], and specificity of 78.4% [0.71 – 0.86] (Figure 1A). For PSE, there was a PPV of 71.0%, NPV of 97.0%, accuracy of 84.0%, sensitivity of 95.9% [0.86 – 1.06], and specificity of 77.0% [0.70 – 0.84] (Figure 1B).

Conclusions: We found that using the EDV encounter data and ICD codes previously validated for epilepsy, when demarcated by the time of entry in the EMR, identify ASyS and PSE with excellent sensitivity and acceptable specificity compared to ground truth. We believe this method of accurately automating EMR phenotyping of post-stroke seizures will reduce the resource utilization burden in developing PSE predictive models from large-scale real-world patient data.

Funding: No funding was received in support of this abstract.