Abstracts

Rationale: Referral of patients with epilepsy (PWE) who develop resistance to antiepileptic drugs (AEDs) to specialist centers is typically delayed by several years, causing substantial disability and deterioration in quality of life for many. Identifying patients at high risk of DRE at the time they receive their first AED has high clinical utility, as early referral can lead to adoption of more personalized, aggressive therapeutic approaches. Here we describe the development of machine-learning models to identify patients who have, or are at high risk of DRE.  Methods: Longitudinal, multichannel, open administrative claims data for >2 million US-based patients from 1 Jan 2006 to 31 Dec 2015 were scanned. Records of PWE were included if the patient received a first AED prescription, and was ≥16 years of age at prescription. From 52,060 records identified, 1,968 patient features on demographics, comorbidities, medications, procedures, and epilepsy and payer status were extracted and used to build models based on 3 techniques: binomial logistic regression (BLR), support vector machines and random forest. Data from 36,442 patients selected randomly were used to train the models and from the remainder to assess the predictive power of the trained models. A benchmark model using only age and sex was built for comparison. To allow for optimal separation between patients with DRE and those with treatment success, DRE was defined as failure of 3, rather than 2 AEDs during the time covered in the dataset. Treatment success was defined as continuous coverage of the AED for ≥12 months after initiation without addition of or switch to another AED (dose changes allowed).  Results: Among 52,060 PWE starting their first AED between 2006 and 2015, 5,554 (10.6%) were classified as having DRE according to study criteria. The most successful model, BLR, achieved an area under the receiver operating characteristic curve (95% CI) of 0.7569 (0.7498, 0.7641), representing a meaningful improvement over the benchmark (AUC 0.664, 95% CI 0.658, 0.671); an AUC of 1.0 indicates a perfectly accurate prediction. The model’s positive and negative predictive values were 0.4480 and 0.8662, respectively.  Conclusions: Numerous studies have identified risk factors for DRE; however, evidence has been inconsistent. Use of large volumes of data, and powerful computational techniques can help overcome limitations of standard clinical prediction models. The predictive model developed here showed potential in identifying patients at high risk of DRE at the time of the first AED prescription. The use of such models can help clinicians identify high-risk patients at first consultation, allowing early referral to specialist care. Ongoing work, including validation of the proxies used for DRE and treatment success, will allow further refinement of the model. Funding: UCB Pharma-sponsored