Abstract
PURPOSE: The methacholine challenge test (MCT) has high sensitivity but relatively low specificity for asthma diagnosis. This study aimed to develop and validate machine learning (ML) models to improve the diagnostic performance of MCT for asthma. METHODS: Data from 1,501 patients with asthma symptoms who underwent MCT between 2015 and 2020 were analyzed. The patients were grouped as either the training (80%, n = 1,265) and test sets (20%, n = 236) depending on the time of referral. The conventional model (provocative concentration that causes a 20% decrease in forced expiratory volume in one second [FEV(1)]; PC(20) ≤ 16 mg/mL) was compared with the prediction models derived from five ML methods: logistic regression, support vector machine, random forest, extreme gradient boosting, and artificial neural network. The area under the receiver operator characteristic curves (AUROC) and area under the precision-recall curves (AUPRC) of each model were compared. The prediction models were further analyzed using different input combinations of FEV(1), forced vital capacity (FVC), and forced expiratory flow at 25%-75% of forced vital capacity (FEF(25%-75%)) values obtained during MCT. RESULTS: In total, 545 patients (36.3%) were diagnosed with asthma. The AUROC of the conventional model was 0.856 (95% confidence interval [CI], 0.852-0.861), and the AUPRC was 0.759 (95% CI, 0.751-0.766). All the five ML prediction models had higher AUROC and AUPRC values than those of the conventional model, and random forest showed both highest AUROC (0.950; 95% CI, 0.948-0.952) and AUROC (0.909; 95% CI, 0.905-0.914) when FEV(1), FVC, and FEF(25%-75%) were included as inputs. CONCLUSIONS: Artificial intelligence-based models showed excellent performance in asthma prediction compared to using PC(20) ≤ 16 mg/mL. The novel technology could be used to enhance the clinical diagnosis of asthma.