INTRODUCTION

Adverse drug reactions (ADRs) are appreciably harmful or unpleasant reactions resulting from an intervention related to the use of a medicinal product. They usually predict hazard from future administrations and warrant prevention, or specific treatment, or alteration of the dosage regimen, or withdrawal of the drug [1]. ADRs are common in clinical settings as they may produce unscheduled hospital admissions, may occur during hospital stay, and may manifest after discharge [2,3]. Around 15% of ADRs are drug hypersensitivity reactions (DHRs) [4], which may be severe and threat patient's life, and are responsible for a huge burden on public healthcare systems. 

Box 1:

no caption available

DHRs may be induced by specific immune mechanisms (drug allergy, DA), with betalactam antibiotics (BLs) being the most frequent triggers [5▪,6,7]. DA reactions can be primarily classified into immediate and nonimmediate reactions (IRs and NIRs, respectively), which correspond to the so-called Type I and Type IV reactions in the classification of Gell and Coombs [8,9]. IRs, appearing within 1–6 h after drug intake, are mediated by specific immunoglobulin E (IgE) antibodies, and the main clinical phenotypes are urticaria, angioedema, bronchospasm, pruritus, and anaphylaxis. NIRs, appearing more than 6 h after drug intake, usually after 24 h, are mainly mediated by T lymphocytes, with a great heterogeneity of clinical phenotypes ranging from urticaria/maculopapular exanthema to more severe reactions such as Stevens-Johnson syndrome/toxic epidermal necrolysis (SJS/TEN), acute generalized exanthematous pustulosis, drug reaction with eosinophilia and systemic symptoms (DRESS), and single-organ reactions such as drug-induced liver disease (DILI), among others [10].

Although a description of diagnostic protocols used in clinical settings to manage DA goes beyond our aim, several tools are available, including both in vivo and in vitro tests [11–13,14▪]; however, controversies and differences between countries exist [15,16]. Independently, a detailed clinical history is crucial to deal with these patients, and helps us to classify/stratify them to elaborate clinical algorithms and implement predictive models for precision medicine [17]. In addition to diagnosis, these models could also help to deal with individuals labelled as allergic even in absence of a clearly established diagnosis, which entails the prescription of alternative medicines that are frequently less efficient and more toxic and expensive [18▪,19]. This is particularly important regarding BLs, the most commonly used antibiotics worldwide, as it is estimated than 15% of the hospital population and 8% of the general population are labelled as allergic to BLs [20]; however, less than 10% of them are confirmed as truly allergic [10].

Progress undertaken over recent years in data treatment and computer sciences have led to the development of different branches of artificial intelligence (AI), such as machine and deep learning (ML and DL, respectively). The most technologically advanced approaches use a diversity of computing systems of interconnected nodes trying to imitate biological neuronal networks, collectively designed as artificial neural networks (ANN), with potential to provide reliable and optimal models for clinical diagnosis, prediction, and precision medicine in human diseases, along with DA (Table 1) [21].

Table 1 - General concepts in artificial intelligence Term Definition Artificial intelligence (AI) Computer science knowledge field that studies and develops computerized mathematical methods which capacitates machines to mimic human intellectual abilities. Algorithm Set of rules or instructions used to solve a particular problem. Model In AI, an algorithm trained to output decisions or interpretations according to certain input data. Machine learning (ML) Development and study of AI algorithms that improve its accuracy gradually learning from a given data set. Deep learning (DL) An advanced branch of ML that mimics by arranging multiple algorithm layers to solve a specific task. Artificial neural network (ANN) A set of connected units or nodes (artificial neurons), linked by edges. Each artificial neuron receives signals from connected neurons, then processes them and sends a signal to other connected neurons. The “signal” is a real number, and the output of each neuron is computed by some nonlinear function of the sum of its inputs, called the activation function. The strength of the signal at each connection is determined by a weight, which adjusts during the learning process.

In this article, we will address some general concepts in AI-assisted diagnosis (AIAD) and the requirements to build IA-based predictive models for DA. Finally, we will describe available predictive models and explore current needs for successfully developing these approaches in DA. We believe the information we provide here could be of interest not only for researchers involved in the study of these pathologies but also for clinicians managing these patients.

GENERAL CONCEPTS IN ARTIFICIAL INTELLIGENCE-ASSISTED DIAGNOSIS

AIAD implies the analysis and process of some input data, such a medical datasets or images, through different computer programs and algorithms to assist clinicians to establish patient clinical phenotypes and to achieve a consistent diagnosis and an appropriate treatment [22▪,23▪]. This approach could help to reduce costs and timing, and to take advantage of the multiple sources of information available [23▪]. Consequently, AIAD provides healthcare professionals with a full interpretation of usable information for them to apply the appropriate treatment.

In its simplest form, AIAD is a binary classification problem in which it is questioned whether a patient has a certain pathology or not. An issue that could be addressed with single classical statistical models such as logistic regression (LR) or hierarchical models like decision or classification trees, still individual models usually underperform in real life diagnosis so there is a need to apply more complex statistic algorithms that learns from the data, such as those used in ML an DL [23▪,24▪▪].

ML is a branch of AI science that involves the development and deployment of algorithms that learn from previous data to make predictions, i.e. generalize, with new data. In other words, if applied to diagnosis, ML models would study multiples times a clinical dataset to find a combination of data, sequential or not, to properly classify patients into at least two clinical categories. Algorithms performance improves gradually applying mathematical programming, and thus selecting the best combination regarding a specific criterion [24▪▪]. Based on this criterion, ML algorithms can be classified into reinforcement, unsupervised, and supervised learning/training. Diagnostic algorithms are mostly based on supervised optimization because models learn from known categories of patients, such as healthy or sick, hence model performance is measured for its accuracy to diagnose patients that are correctly diagnosed manually by clinicians.

Supervised learning has a major downside, overfitting, this means that in the successive cycles of training the ML algorithm could learn to correctly classify the patients included in the training but underperforms classifying new patients of an unseen dataset [24▪▪,25,26]. To resolve this issue, the dataset collected to train the model is partitioned into at least two sets, one for training and another for testing the model, so it is necessary to gather the largest possible amount of curated data. Some remarkable examples of supervised ML algorithms are Random Forest, Naive Bayes or support-vector machines [24▪▪].

In some cases, automated diagnosis requires the use of more complex algorithms to reach an accuracy appropriate for its use in clinical practice. This group of statistic models based on multiple layers of ANN are known as DL, for supervised learning each one of the layers transforms the data to establish connection patterns that label the patients in the last layer (output), in such a way that the data/signal travels from the input layer to the output layer not necessarily linear, since it is inspired by biological neuron networks, signal that can travel backwards [24▪▪].

BUILDING AN ARTIFICIAL INTELLIGENC-BASED PREDICTIVE MODEL IN DRUG ALLERGY

A theoretical approach to build an IAAD tool could be developed in two general phases (Fig. 1). The first phase includes curating and preparing the dataset, and studying the information contained in the database. It would seem appropriate to use all available variables; however, some of them may correspond with diagnostic tests or clinician's interpretations that occur after the time window in which the algorithm is to be applied. Ideally, to reduce DA diagnosis costs and reduce human resources and intervention, automated diagnosis algorithm should learn to classify suspected patients based on clinical history and hypersensitivity reaction recorded variables. Next, once the dataset is curated and delimited, it should be randomly divided into training (∼80%) and testing (∼20%) subsets preserving the ratio of TRUE/FALSE BL allergy patients, the 80/20 split is indicative and could change based on dataset variability and size; this step can be done e.g. by creating a Data Partition function in caret R package [25]. Before proceeding with the training, additional optional steps can be applied to the dataset: missing data imputation and oversampling; these processes allow to complete the missing information in the dataset with mean/mode/median imputation, and to balance the ratio of patients with artificial patients created with different statistical algorithms such as ADASYN, SMOTE and RACOG, among others [26].

FIGURE 1:

Graphical summary of the development and implementation of a ML model to diagnose drug allergy patients. Left panel represents data collection and curation. Right panel outlines the processes involved in the development of the classification algorithm. ML, machine learning.

In the second phase, different ML models would be trained to diagnose patients. A good practice is to apply sequentially different supervised ML algorithms increasing model complexity each time. For example, a tree-based algorithm approach could be used, beginning with decision trees [27], followed by Random Forest [25], and ending with Light Gradient Boosting Machine [28]. Finally, from training each algorithm an optimum is obtained after a predefined number of iterations, which is measured with certain performance metrics such as the area under the curve (AUC). Then, each implementation performance is tested against the test set, ML models usually underperforms in the evaluations against the test dataset as the learning process tends to overfit the algorithm and its parameters to the training data. To overcome the overfitting problem, ML algorithms include different strategies, like validation datasets (hold-out), cross-validation, early stopping, feature selection or data augmentation. Moreover, feature selection strategy also reduces the number of variables needed to predict a diagnosis, therefore facilitating future application in real life situations, as well as training time and computer resources demand. In this procedure, variables could be selected by a combination of parameters, for example how much information is gained by splitting the data by a given variable (gain). And feature selection could also be done employing wrapper (sequential forward or backward selection), embedded (e.g. LASSO) or filter methods (e.g. recursive feature elimination); caret R package includes algorithms with built-in feature selection [25]. At this point, one of the models is selected according to its performance in diagnosing DA, this is measured with diverse well known classification metrics, e.g. sensitivity, specificity, accuracy, AUC, etc. Depending on the application, some metrics will be prioritized. In DA diagnosis, high sensitivity models could be developed to avoid allergic reactions in high-risk individuals, and high specificity models could be developed to avoid further studies or treatments in low risk suspected patients.

Finally, the selected predictive model should be used to predict DA in new suspected cases in a prospective study to verify its results and even to retrain the algorithm adding novel information into the model, which would improve its future performance in real life situations.

PREDICTIVE MODELS IN DRUG ALLERGY

Classical statistical analyses of complex medical datasets have shown their drawback to manage independent nonlinear relationships between variables [29], a limitation that can be outdone by AI-related approaches. Despite these approaches currently revealing their potential to benefit healthcare systems by improving patients’ stratification, diagnosis and therapy in human diseases [30,31], including allergy [32], there is a limited number of studies on DA.

Risk stratification of patients is providing considerable progress in DA, particularly in BL-allergy, as this is a consistent method to identify those individuals at low-risk in which skin test may be avoidable and direct oral challenge to be considered as safe, with the subsequent savings in time and costs [12,33]. Stratification of BL-allergic patients can specially benefit from the use of predictive models. In an Australian retrospective multicentric study dealing with the identification of low- and high-risk BL-allergic patients, using different LR analysis, low-risk patients were identified as those showing a history of penicillin-associated rash more than 1 year ago without developing angioedema, mucosal ulceration, or systemic involvement [34]. However, in addition to the lack of skin test data for all patients and the heterogeneous testing strategies used, this model was not subjected to internal and external confirmation.

In another study trying to construct a predictive model, both a retrospective and prospective populations were employed [35]. Clinical data from the retrospective population (around 2000 patients) were used to build a LR-based predictive model and a decision tree using rpart. These two approaches were further validated externally in the prospective population (200 patients). However, the LR model presented low sensitivity and a number of predictors or combination of them that would be probably difficult to use in clinical settings. Furthermore, the decision tree was also unable to accurately predict BL-allergy [35]. By combining univariate analysis and multiple LR, a combined model including a family history of drug allergy, anaphylaxis, any atopic disease other than asthma, a reaction interval of 0–6 h showed a sensitivity of 77.8%, and a negative predictive value of 94.3% in predicting BL-allergy in children, with these values being further replicated in a second cohort [36].

Another attempt to design a decision tool is represented by the user-friendly PEN-FAST questionnaire, a penicillin allergy clinical decision rule that used five clinical history questions to predict patient's risk to have a true BL-allergy [37]. This questionnaire has allowed a correct delabelling of low-risk patients in other populations [38▪▪,39▪▪,40,41]; however, it did not help to identify low-risk penicillin-allergy in a Canadian paediatric prospective multicentred cohort [42▪▪]. Future efforts to apply this strategy should increase the number of questions or items included to improve its performance as a classification algorithm. An example is the PEN-FAST+ score, which has been developed to increase its accuracy for immediate skin and delayed maculopapular exanthema reactions, as the previous clinical decision tool misclassified 28.6% of patients with IR and 38.4% of patients with NIR [43].

As far as we known, the first available study designing an AI-based model in BL-allergy was that by Moreno et al., which used more complex computing resources to enhance the predictive model through ANN. Handling a retrospective and a prospective population of BL-allergic patients, they compared a LR model with an ANN approach [44]. In these two series of patients, ANN reached a sensitivity of 89.5% and 81.1%, a specificity of 86.1% and 97.9%, a positive predictive value of 82.1% and 91.1%, and a negative predictive value of 92.1% and 95.2%, respectively. Moreover, ANN's performance was far superior to that of LR, whose best performance reached a sensitivity of 31.9% and a specificity of 98.8%. [44]. Interestingly, some studies have used DL algorithms with another target, instead of developing models for just classifying patients, they employed natural language processing DL algorithms to detect true drug, including BL-allergy reactions from electronic health records [45,46].

Concerning NIR-DHRs, in a recent study in DILI patients, most of them reacting to penicillin, a ML approach has revealed that a previous history of DA is an independent predictor of fatal outcome, and allowed to develop a useful tool based on their validated model for risk stratification of these patients [47▪▪]. Moreover, recently, the use of electronic health records has shown utility to build a time-series DL model to predict DILI in patients taking angiotensin receptor blockers, with a performance depending on the specific drug; however, this model has not been externally validated [48▪▪]. Other models using LR have dealt with the identification of key factors predicting mortality in patients with SJS/TEN [49▪▪], although not all have included a validation population [50▪▪].

In addition to be used directly in predictive models for DHRs, AI-derived models may be also used in pharmacokinetics, because serum drug levels may affect the development of some reactions, as it has been suggested for vancomycin-induced DRESS [51]. Thus, a DL-based approach has shown its utility in vancomycin treatment monitoring to deal with patients suffering critical illnesses, as demonstrated by its superior performance compared to other population pharmacokinetic models [52▪▪]. Finally, another ML model has recently achieved the highest precision in predicting NIR using IFN-γ ELISpot and clinical variables [53▪▪].

NEEDS AND PERSPECTIVES

Future AI-based predictive models in DA should begin with extensive data collection and curation. As mentioned, past efforts failed to build a predictive model able to perform well in populations other than the original training/testing population. This is because, while AI science is rapidly developing new algorithms and strategies, DA datasets still lack diversity in different variables (age, gender, symptoms and ethnicity, among others) in order to generalize the performance of AI models. This mainly occurs because most of implementations are based solely on the data collected by a few or even one allergy unit, biasing the information by the formal process of diagnosis of each unit and by the subject population. For this reason, it is necessary to involve multiple centres that would provide enough diverse data that would enable the development of high performing classification algorithms.

Acknowledgements

None.

Financial support and sponsorship

None.

Conflicts of interest

There are no conflicts of interest.

REFERENCES AND RECOMMENDED READING

Papers of particular interest, published within the annual period of review, have been highlighted as:

▪ of special interest

▪▪ of outstanding interest

REFERENCES 1. Aronson JK, Ferner RE. Clarification of terminology in drug safety. Drug Saf 2005; 28:851–870. 2. Lazarou J, Pomeranz BH, Corey PN. Incidence of adverse drug reactions in hospitalized patients: a meta-analysis of prospective studies. JAMA 1998; 279:1200–1205. 3. Pirmohamed M, James S, Meakin S, et al. Adverse drug reactions as cause of admission to hospital: prospective analysis of 18 820 patients. BMJ 2004; 329:15–19. 4. Thong BY, Tan TC. Epidemiology and risk factors for drug allergy. Br J Clin Pharmacol 2011; 71:684–700. 5▪. Doña I, Torres MJ, Celik G, et al. Changing patterns in the epidemiology of drug allergy. Allergy 2024; 79:613–628. 6. Salas-Casinello M, Saenz-de Santa Maria R, Lopez-Sanchez JD, et al. Different patterns of response in hypersensitivity reactions to arylpropionic acid derivatives. J Allergy Clin Immunol Pract 2023; 11:3715–3723. 7. Minaldi E, Phillips EJ, Norton A. Immediate and delayed hypersensitivity reactions to beta-lactam antibiotics. Clin Rev Allergy Immunol 2022; 62:449–462. 8. Demoly P, Adkinson NF, Brockow K, et al. International consensus on drug allergy. Allergy 2014; 69:420–437. 9. Pichler WJ. Immune pathomechanism and classification of drug hypersensitivity. Allergy 2019; 74:1457–1471. 10. Castells M, Khan DA, Phillips EJ. Penicillin allergy. N Engl J Med 2019; 381:2338–2351. 11. Doña I, Romano A, Torres MJ. Algorithm for betalactam allergy diagnosis. Allergy 2019; 74:1817–1819. 12. Romano A, Atanaskovic-Markovic M, Barbaud A, et al. Towards a more precise diagnosis of hypersensitivity to beta-lactams – an EAACI position paper. Allergy 2020; 75:1300–1315. 13. Barbaud A, Castagna J, Soria A. Skin tests in the work-up of cutaneous adverse drug reactions: a review and update. Contact Dermatitis 2022; 86:344–356. 14▪. Barbaud A, Garvey LH, Torres M, et al. EAACI/ENDA position paper on drug provocation testing. Allergy 2024; 79:565–579. 15. Torres MJ, Adkinson NF Jr, Caubet JC, et al. Controversies in drug allergy: beta-lactam hypersensitivity testing. J Allergy Clin Immunol Pract 2019; 7:40–45. 16. Torres MJ, Romano A, Celik G, et al. Approach to the diagnosis of drug hypersensitivity reactions: similarities and differences between Europe and North America. Clin Transl Allergy 2017; 7:7. 17. Plager J, Judd A, Blumenthal K. Role of clinical history in beta-lactam hypersensitivity. Curr Opin Allergy Clin Immunol 2021; 21:320–326. 18▪. Saenz de Santa Maria R, Bogas G, Labella M, et al. Approach for delabeling beta-lactam allergy in children. Front Allergy 2023; 4:1298335. 19. Doña I, Labella M, Bogas G, et al. Antibiotic allergy de-labeling: a pathway against antibiotic resistance. Antibiotics (Basel) 2022; 11:1055. 20. Harandian F, Pham D, Ben-Shoshan M. Positive penicillin allergy testing results: a systematic review and meta-analysis of papers published from 2010 through. Postgrad Med 2016; 128:557–562. 21. Blackley SV, Salem A, Zhou L. Deep learning for detection of drug hypersensitivity reactions. J Allergy Clin Immunol 2023; 152:350–352. 22▪. Wimpfheimer O, Kimmel Y. Artificial intelligence in medical imaging: an overview of a decade of experience. Isr Med Assoc J 2024; 26:122–125. 23▪. Krishnan G, Singh S, Pathania M, et al. Artificial intelligence in clinical medicine: catalyzing a sustainable global healthcare paradigm. Front Artif Intell 2023; 6:1227091. 24▪▪. Terranova N, Venkatakrishnan K. Machine learning in modeling disease trajectory and treatment outcomes: an emerging enabler for model-informed precision medicine. Clin Pharmacol Ther 2024; 115:720–726. 25. Kuhn M. Building predictive models in R using the caret package. J Statist Softw 2008; 28:1–26. 26. Cordón IGS, Fernández A, Herrera F. Imbalance: oversampling algorithms for imbalanced classification in R. Knowledge-Based Syst 2018; 161:329–341. 27. Terry Therneau BA, Brian Ripley. Rpart: recursive partitioning for classification, regression and survival trees. Available at: https://cran.r-project.org/web/packages/rpart/index.html. 2013. 28. Guolin Ke QM, Thomas Finely, Taifeng Wang, et al. LightGBM: a highly efficient gradient boosting decision tree. Proceedings of the 31st International Conference on Neural Information Processing Systems Curran Associates Inc 2017; 3149–3157. 29. Hashimoto DA, Rosman G, Rus D, Meireles OR. Artificial intelligence in surgery: promises and perils. Ann Surg 2018; 268:70–76. 30. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019; 25:44–56. 31. Esteva A, Robicquet A, Ramsundar B, et al. A guide to deep learning in healthcare. Nat Med 2019; 25:24–29. 32. MacMath D, Chen M, Khoury P. Artificial intelligence: exploring the future of innovation in allergy immunology. Curr Allergy Asthma Rep 2023; 23:351–362. 33. Banks TA, Tucker M, Macy E. Evaluating penicillin allergies without skin testing. Curr Allergy Asthma Rep 2019; 19:27. 34. Stevenson B, Trevenen M, Klinken E, et al. Multicenter Australian study to determine criteria for low- and high-risk penicillin testing in outpatients. J Allergy Clin Immunol Pract 2020; 8:681–689.e3. 35. Chiriac AM, Wang Y, Schrijvers R, et al. Designing predictive models for beta-lactam allergy using the drug allergy and hypersensitivity database. J Allergy Clin Immunol Pract 2018; 6:139–48e2. 36. Demirhan A, Yildirim DD, Arikoglu T, et al. A combined risk modeling strategy for clinical prediction of beta-lactam allergies in children. Allergy Asthma Proc 2021; 42:e159–e166. 37. Trubiano JA, Vogrin S, Chua KYL, et al. Development and validation of a penicillin allergy clinical decision rule. JAMA Intern Med 2020; 180:745–752. 38▪▪. Copaescu AM, Vogrin S, James F, et al. Efficacy of a clinical decision rule to enable direct oral challenge in patients with low-risk penicillin allergy: the PALACE Randomized Clinical Trial. JAMA Intern Med 2023; 183:944–952. 39▪▪. Su C, Belmont A, Liao J, et al. Evaluating the PEN-FAST clinical decision-making tool to enhance penicillin allergy delabeling. JAMA Intern Med 2023; 183:883–885. 40. Copaescu AM, James F, Vogrin S, et al. Use of a penicillin allergy clinical decision rule to enable direct oral penicillin provocation: an international multicentre randomised control trial in an adult population (PALACE): study protocol. BMJ Open 2022; 12:e063784. 41. Piotin A, Godet J, Trubiano JA, et al. Predictive factors of amoxicillin immediate hypersensitivity and validation of PEN-FAST clinical decision rule. Ann Allergy Asthma Immunol 2022; 128:27–32. 42▪▪. Copaescu AM, Vogrin S, Shand G, et al. Validation of the PEN-FAST score in a pediatric population. JAMA Netw Open 2022; 5:e2233703. 43. Castagna J, Chasset F, Autegarden JE, et al. Assessing delayed penicillin hypersensitivity using the PENFAST+ score. Front Allergy 2023; 4:1302567. 44. Moreno EM, Moreno V, Laffond E, et al. Usefulness of an artificial neural network in the prediction of beta-lactam allergy. J Allergy Clin Immunol Pract 2020; 8:2974–3820. 45. Inglis JM, Bacchi S, Troelnikov A, et al. Automation of penicillin adverse drug reaction categorisation and risk stratification with machine learning natural language processing. Int J Med Inform 2021; 156:104611. 46. Lo YC, Varghese S, Blackley S, et al. Reconciling allergy information in the electronic health record after a drug challenge using natural language processing. Front Allergy 2022; 3:904923. 47▪▪. Niu H, Solis-Munoz P, Garcia-Cortes M, et al. Prior drug allergies are associated with worse outcome in patients with idiosyncratic drug-induced liver injury: a machine learning approach for risk stratification. Pharmacol Res 2024; 199:107030. 48▪▪. Heo S, Yu JY, Kang EA, et al. Development and verification of time-series deep learning for drug-induced liver injury detection in patients taking angiotensin ii receptor blockers: a multicenter distributed research network approach. Healthc Inform Res 2023; 29:246–255. 49▪▪. Hama N, Sunaga Y, Ochiai H, et al. Development and validation of a novel score to predict mortality in stevens-johnson syndrome and toxic epidermal necrolysis: CRISTEN. J Allergy Clin Immunol Pract 2023; 11:3161–3800. 50▪▪. Chuenboonngarm N, Puaratanaarunkon T, Washrawirul C, et al. Key factors predicting the in-hospital mortality of patients with severe cutaneous adverse reactions in Thailand. J Eur Acad Dermatol Venereol 2023; 37:1881–1890. 51. Mitsuno R, Nakayama T, Uchiyama K, et al. Hemodialysis treatment of vancomycin-induced drug reaction with eosinophilia and systemic symptoms/drug-induced hypersensitivity syndrome in a patient undergoing peritoneal dialysis. CEN Case Rep 2024; doi: 10.1007/s13730-023-00847-x. [Online ahead of print]. 52▪▪. Kim D, Choi HS, Lee D, et al. A deep learning-based approach for prediction of vancomycin treatment monitoring: retrospective study among patients with critical illness. JMIR Form Res 2024; 8:e45202. 53▪▪. Chongpison Y, Sriswasdi S, Buranapraditkun S, et al. IFN-gamma ELISpot-enabled machine learning for culprit drug identification in nonimmediate drug hypersensitivity. J Allergy Clin Immunol 2024; 153:193–202.