Background: Nonsteroidal anti-inflammatory drugs (NSAIDs) are the main triggers of drug hypersensitivity reactions (HSRs). NSAID-induced HSRs constitute a significant health problem owing to their prevalence, phenotypes, mechanisms of action, and complex diagnosis. However, in around 60% of individuals attending for suggestive NSAID-induced HSRs, this diagnosis is ruled out. Patients labeled as being hypersensitive to NSAIDs unnecessarily avoid NSAIDs, leading to increased waiting lists, diagnostic delay, and associated costs. Objective: To develop a machine learning (ML)–based model for discriminating NSAID-hypersensitive patients from non–NSAID-hypersensitive individuals by assessing populations with suspected NSAID-induced HSRs. Methods: We recruited a retrospective population and a prospective population of individuals attending the Allergy Unit of Malaga Regional University Hospital, Malaga, Spain for suggestive NSAID-induced HSRs in whom a diagnosis had been confirmed. One logistic regression analysis and 6 ML-based models were developed using retrospective data, and the most efficient was applied to the prospective population. In addition, 3 prospective populations from Madrid, Barcelona, and Salamanca were included for external validation. Results: All the models classified at least 85% of individuals correctly, and, considering discrimination by chance, agreement was almost perfect for all of them (k>0.81). However, the light gradient-boosting machine (LGBM) model showed the highest sensitivity (99%), accuracy (97%), and k value (0.94). The final validated LGBM model achieved 91.76% accuracy, a 95.19% area under the curve, and a k of 0.83. Accuracy was >95% in all prospective populations. Conclusion: Our LGBM model efficiently differentiated NSAID-hypersensitive patients from individuals who can safely receive NSAIDs, despite suggestive NSAID-induced HSRs. Such a model could easily be incorporated into clinical settings, thus improving diagnosis, reducing waiting lists, and optimizing health care resources.
María et al. (Mon,) studied this question.