Abstracts

Rationale: The purpose of this study was use a combination of clinical intuition and data driven approaches to derive simple clinical models to predict which epilepsy patients are at risk of developing psychiatric adverse effects from levetiracetam. Methods: We first used a modified Delphi method and a literature review to identify candidate risk factors for model inclusion. We then extracted incident cases of epilepsy in The Health Improvement Network database, inclusive years 2000-2012, using a five-year washout. Index date was that on which patients received their first prescription code for levetiracetam and follow-up lasted two-years. We defined the outcome of interest as the presence of a code for any psychiatric symptom or disorder reached through consensus. We used univariable and multiple logistic regression to derive the prediction model. As a sensitivity analysis, we performed the same analyses but only on those without a history of a psychiatric sign, symptom, or disorder to infer the predictive value of these variables on incident psychiatric complaints following the first levetiracetam prescription. Model calibration and discrimination were measured using the Brier score and area under the receiver operating characteristic curve (AUC) and k=5 fold cross validation was used to evaluate generalizability. Results: We identified 9595 patients with incident epilepsy of whom 1173 (14%) received a first-ever prescription for levetiracetam. A total of 14% (165/1173) were coded with a psychiatric symptom or disorder within two-years of index prescription. Female sex, greater social deprivation, a pre-prescription history of a code for depression or anxiety, and recreational drug use were each independently associated with increased odds of reporting a psychiatric sign or symptom after their initial recorded levetiracetam prescription. The final prediction model, incorporating sex, history of depression, history of anxiety, and recreational drug use, performed well after stratified k=5 fold cross-validation (AUC 0.68; 95% confidence interval [95%CI] 0.58-0.79) with no concerns about poor model calibration or goodness of fit. The number of risk factors present contributed to an incremental gradient in risk, with probabilities of incurring a psychiatric sign, symptom, or disorder following incident prescription increasing from 8% for zero risk factors, 11-17% for one, 17-31% for two, 30-42% for three, and 49% when all risk factors are present. In the sensitivity analysis, we identified 710/1173 (61%) patients without a history of a psychiatric code. In this model, increasing age, female sex, recreational drug use, increasing number of concurrent AEDs, and increasing levetiracetam dose were all predictive of a presumed incident psychiatric sign, symptom, or disorder following the first levetiracetam prescription. Again, this model performed well after stratified k=5 fold cross-validation (AUC 0.74; 95%CI 0.58-0.90) with no concerns about poor model calibration or goodness of fit. Conclusions: We derived two simple prediction models for the risk of a psychiatric event following a presumed first prescription for levetiracetam. Of all identified risk factors, depression and recreational drug use appear to confer the highest risk for those with a history of psychiatric signs or symptoms. There was also a graded risk according to the number of risk factors present. For those without a history of psychiatric signs or symptoms, increasing age, female sex, levetiracetam dose, and AED polytherapy were all associated with increased risk. Validation in external populations will be beneficial in confirming the general utility of this model in clinical practice. Funding: Not applicable