IFN-γ ELISpot-enabled machine learning for culprit drug identification in nonimmediate drug hypersensitivity
Yuda Chongpison1, Sira Sriswasdi2, Supranee Buranapraditkun3
1Biostatistics Excellence Centre, Research Affairs, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand; Skin and Allergy Research Unit, Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand.
Background:
Diagnosing drug-induced allergy, especially nonimmediate phenotypes, is challenging. Incorrect classifications have unwanted consequences.
Objective:
We sought to evaluate the diagnostic utility of IFN-γ ELISpot and clinical parameters in predicting drug-induced nonimmediate hypersensitivity using machine learning.
Methods:
The study recruited 393 patients. A positive patch test or drug provocation test (DPT) was used to define positive drug hypersensitivity. Various clinical factors were considered in developing random forest (RF) and logistic regression (LR) models. Performances were compared against the IFN-γ ELISpot-only model.
Results:
Among the 102 patients who had 164 DPTs, most patients had severe cutaneous adverse reactions (35/102, 34.3%) and maculopapular exanthems (33/102, 32.4%). Common suspected drugs were antituberculosis drugs (46/164, 28.1%) and β-lactams (42/164, 25.6%). Mean (SD) age of patients with DPT was 52.7 (20.8) years. IFN-γ ELISpot, fixed drug eruption, Naranjo categories, and nonsteroidal anti-inflammatory drugs were the most important features in all developed models. The RF and LR models had higher discriminating abilities. An IFN-γ ELISpot cutoff value of 16.0 spot-forming cells/106 PBMCs achieved 94.8% specificity and 57.1% sensitivity. Depending on clinical needs, optimal cutoff values for RF and LR models can be chosen to achieve either high specificity (0.41 for 96.1% specificity and 0.52 for 97.4% specificity, respectively) or high sensitivity (0.26 for 78.6% sensitivity and 0.37 for 71.4% sensitivity, respectively).
Conclusions:
IFN-γ ELISpot assay was valuable in identifying culprit drugs, whether used individually or incorporated in a prediction model. Performances of RF and LR models were comparable. Additional test datasets with DPT would be helpful to validate the model further.
