Identifying drug-resistant epilepsy using Swedish National Patient Registry data

Approximately 30% of individuals with epilepsy are drug-resistant, suffering low quality of life and should be evaluated for other treatment modalities. Drug-resistant epilepsy has no specific diagnostic code, making identification in national healthcare registers challenging and limiting epidemiological research based on register data.

To define our cohort, we used 1320 individuals from the Swedish National Epilepsy Surgery Register (SNESUR) as index individuals. Each index individual was matched with up to 17 individuals with epilepsy of similar age and diagnosis year in the Swedish National Patient Register. After excluding all SNESUR individuals, the study cohort consisted of 21 393 individuals, a cohort broadly representative of the general epilepsy population. Data on diagnoses, procedures, healthcare contacts, and dispensed antiseizure medications were obtained from national registers. A distance-based ranking algorithm was developed and validated against medical records. Three machine-learning classifiers were then trained and tested using a proxy of the International League Against Epilepsy definition of drug-resistant epilepsy.

The ranking algorithm showed high performance for identifying drug-resistant epilepsy, with a positive predictive value (PPV) of 86.2% (95% CI: 0.81-0.90). Out of the three machine-learning classifiers, Gradient Boosting Machine achieved the highest F1-score of 0.762 (along with PPV = 0.781, sensitivity = 0.744, specificity = 0.820, area under the precision-recall curve = 0.858). The most significant variables found were "total visits to outpatient clinic due to epilepsy," "age of onset of epilepsy," "total admissions due to epilepsy," and "total days of admission due to epilepsy."

Both the ranking algorithm and machine-learning approaches showed good performance in register-based phenotyping of drug-resistant epilepsy using routinely collected register data, supporting their use in epidemiologic research.