D3.25 - Machine learning models to differentiate cross-reactive and selective-hypersensitive patients to nonsteroidal anti-inflammatory drugs hypersensitivity

Feasibility and applicability of machine learning (ML)-poweredautomated stratification of hypersensitivity reactions to nonsteroidal anti-inflammatory drugs (NSAID-HSRs) have been recently successfully explored. Itsimplementation may reduce the NSAID-HSR burden on healthcare systems,speeding up diagnosis and improving quality of life of finally de-labelled healthypatients. In addition to be accurate in differentiating hypersensitive from non-hypersensitive individuals, automated diagnosis of NSAID-HSRs should alsodiscriminate cross-reactive from selective patients, whose management in clinicalsettings is complex and time/resource-consuming, and usually requires non-risk-freetests for patients.

Using clinical history data, a nested approach was employed to classifysubjects with a suggestive history of NSAID-HSRs in three categories: cross-reactiveand selective-hypersensitive patients, and non-hypersensitive individuals. In theparent layer, a gradient boosting algorithm (LightGBM), was used to generate amodel classifying subjects in NSAID-hypersensitive patients and non-hypersensitiveindividuals. Next, in the nested layer, a second LightGBM model was trained todiscern between cross- and selective-hypersensitive focusing on those subjectspreviously classified as patients. LightGBM models were trained using twopopulations: retrospective (N=959) and prospective (N=139). To avoid overfitting,data partitioning of both populations was used, leading to three subsets: validation,training and testing. Lastly, an analysis of variable importance was conducted.

In the testing subset, the first LightGBM model achieved an accuracy of98.5%, with 100% sensitivity and 96.8% specificity. Therefore, only 3% of non-hypersensitive individuals were identified as NSAID-hypersensitive patients. Thefinal LightGBM model showed 91.5% accuracy, 90.8% sensitivity, and 92.3%specificity, allowing to correctly classify 92.3% of cross-reactive and 90.8% ofselective-hypersensitive patients in the testing subset. In both prediction models, thevariables with more impact, assessed by gain of accuracy, were: the number ofreported HSRs, age, interval between NSAID-intake and reaction onset, the numberof NSAIDs involved, and development of angioedema.

ML-based methods using clinical history data may be a useful tool forNSAID-HSR diagnosis. They may assist clinicians for personalised patientmanagement, and to prevent inappropriate NSAID-HSR labels that lead individualsto unnecessarily avoid medications they may need now or throughout their lives.