Idegård André, Larsson David
Department of Clinical Neuroscience, Institute of Neuroscience and Physiology, Sahlgrenska Academy, Gothenburg University, Gothenburg, Sweden.
Wallenberg Center of Molecular and Translational Medicine, Gothenburg University, Gothenburg, Sweden.
PLoS One. 2025 Aug 12;20(8):e0329012. doi: 10.1371/journal.pone.0329012. eCollection 2025.
Healthcare administrative data often rely on the International Classification of Diseases (ICD) system, which lacks specific codes to identify etiological subgroups of epilepsy. Combining indicators for epilepsy and potential etiologies is possible, but such approaches require validation. This study aimed to validate methods for identifying poststroke epilepsy (PSE) in Swedish administrative data.
The algorithms were based on combinations of ICD-10 codes for stroke and seizures, with some also incorporating antiseizure medication prescriptions. We focused on positive predictive values (PPVs), using medical records as the reference standard. We identified individuals in the National Patient Register with a primary inpatient diagnostic code for stroke (I61 or I63) during 2005-2010 and a first-ever seizure-related code (G40, G41, or R56.8), occurring more than seven days post-stroke. To facilitate access to medical records, only patients who were deceased at data extraction (Jan 16, 2021) were eligible. A nationwide random sample of 500 patients was selected, with the intended sample for medical record review being 250. Medical records were reviewed before processing the administrative data.
Records were obtained for 321 patients (median age 78; 56% males), with no significant differences in characteristics between those included and the rest of the sample. Across different algorithms, PPVs ranged from 84.1% (95% CI: 79.2-88.3) to 92.5% (95% CI: 87.3-96.1). Relative coverage ranged from 60% to 89% compared to the most inclusive algorithm.
Our findings demonstrate the potential of administrative data to reliably identify PSE cases, supporting the use of these algorithms for large-scale studies of treatment and outcomes. Stricter algorithms, limited to G40 codes for epilepsy or requiring ASM prescriptions, improve accuracy but at the cost of missing more cases. Limitations include the inability to calculate sensitivity due to study design, and the need for local validation before use in other healthcare systems.
医疗保健管理数据通常依赖国际疾病分类（ICD）系统，该系统缺乏用于识别癫痫病因亚组的特定代码。将癫痫指标与潜在病因相结合是可行的，但此类方法需要验证。本研究旨在验证瑞典管理数据中识别卒中后癫痫（PSE）的方法。
算法基于卒中与癫痫发作的ICD - 10代码组合，部分算法还纳入了抗癫痫药物处方。我们以病历作为参考标准，重点关注阳性预测值（PPV）。我们在国家患者登记册中识别出2005 - 2010年期间有卒中主要住院诊断代码（I61或I63）且首次出现与癫痫发作相关代码（G40、G41或R56.8）且在卒中后七天以上发生的个体。为便于获取病历，仅在数据提取时（2021年1月16日）已死亡的患者符合条件。从全国范围内随机抽取了500名患者，计划用于病历审查的样本为250名。在处理管理数据之前先审查病历。
获取了321名患者的病历（中位年龄78岁；56%为男性），纳入患者与其余样本在特征上无显著差异。在不同算法中，PPV范围为84.1%（95%置信区间：79.2 - 88.3）至92.5%（95%置信区间：87.3 - 96.1）。与包容性最强的算法相比，相对覆盖率范围为60%至89%。
我们的研究结果表明管理数据有潜力可靠地识别PSE病例，支持将这些算法用于治疗和结局的大规模研究。更严格的算法，限于癫痫的G40代码或需要抗癫痫药物处方，可提高准确性，但代价是遗漏更多病例。局限性包括由于研究设计无法计算敏感性，以及在用于其他医疗系统之前需要进行本地验证。
