Can nicotine replacement therapy be personalized? A statistical learning analysis

Abstract

BackgroundTechnology has made automated care personalization practical, but useful personalization requires information about systematic differences between individuals in the effectiveness of different interventions. Here, we used observational data to search for differences in smoking cessation treatment outcomes associated with interactions between participant characteristics and different types and doses of nicotine replacement therapy (NRT). MethodsWe analyzed 33,077 enrollments in a large primary care smoking cessation program in Ontario, Canada. We considered 10 types and combinations of NRT, as well as the provided daily dose of nicotine. We used ridge regression to fit one main effects model and one model including all possible interactions between these measures and a range of demographic and health variables. We then compared the predictive accuracy of these models in a held-out 25 % testing subset using areas under the receiver operating characteristic curve (AUROC) and the integrated discrimination improvement index (IDI). We used random forest multiple imputation to address missing data. ResultsThe model including main effects only modestly predicted quit success at 6 months (AUROC = 0.646, 95 % CI = 0.631, 0.660). The final model with all interactions had essentially identical performance (AUROC = 0.640, 95 % CI = 0.626, 0.654; IDI = −0.0066). ConclusionWe found no evidence of meaningful interactions between treatment outcomes and participants' characteristics, NRT type, or NRT dose. Although data are observational, these findings suggest that the effectiveness of different types and doses of NRT do not vary substantially with participant characteristics. Personalization based on the overall likelihood of quit success, or using genetic or other biological data, remains possible.