Abstracts

Rationale:
The standardized mortality ratio (SMR) in epilepsy is 2-3 times higher than the general population and may be up to 4-6-fold higher over the first four years following diagnosis.  Patients desire counseling on premature mortality but clinically informed data-driven models are lacking.
Method:
All patients meeting a case definition for epilepsy in the Acceptable Mortality Recording in The Health Improvement Network database (THIN; 1986-2012) were included. The exposure was a first-ever incident diagnosis of epilepsy using a five-year washout. A modified Delphi process identified 30 potential risk factors. Machine learning methods were used to model early (within 4 years of epilepsy diagnosis) and late (4 years or more from diagnosis) death.
Results:
We identified 12,578 presumed incident cases of epilepsy from 11,194,182 patients registered in THIN. Patients who died were older, more likely to have received an enzyme-inducing antiseizure medication (ASM) at baseline, and more likely to have cardiovascular disease, brain cancer, cirrhosis, chronic obstructive pulmonary disease, dementia, and chronic renal failure. Gaussian Naïve Bayes, Random Forest, linear kernel support vector machines, and logistic regression with parameter regularization classifiers all performed comparably well following stratified 5-fold cross-validation, though logistic regression with parameter regularisation and linear kernel SVM had the highest AUCs for predicting early (0.82; 95%CI 0.79-0.85 and 0.82; 95%CI 0.79-0.85 respectively) and late (0.80; 95%CI 0.74-0.86 and 0.80; 95%CI 0.75-0.85 respectively) death. Both models performed well with excellent Brier scores for early and late death (Table 1). Logistic regression with parameter regularization was well calibrated (Figure 1) and produced clinically intuitive odds ratios (ORs). Using an optimal threshold of 0.13, the sensitivity and specificity for predicting early death was 0.77 and 0.71 respectively for this model. Evaluation of the logistic regression with parameter regularization models revealed treatment patterns acted as significant risk factors. The odds of death were elevated for when no ASM was prescribed at baseline  (odds ratio [OR] for early death 1.44; 95% 1.19-1.75) and following receipt of enzyme-inducing ASM at baseline (OR for late death 1.26; 95%CI 1.04-1.54). Baseline ASM polytherapy was protective for early death (OR 0.65 [95%CI 0.46-0.94]).
Conclusion:
Clinically informed models can be used to predict early and late mortality in epilepsy. The derived models had moderate to high accuracy with evidence of generalisability. Treatment-related risk factors, such as delayed ASM prescription and long-term use of enzyme-inducing ASMs, in addition to medical and social variables, were important predictors.
Funding:
:N/A