001129 - From Symptoms to MBPT: A Machine Learning Approach to Asthma Diagnosis

Accurate asthma diagnosis can be difficult using only patient demographics and symptoms. Integrating clinical data, such as blood tests, fractional exhaled nitric oxide (FeNO), sputum analysis, pulmonary function tests (PFT), and methacholine bronchial provocation tests (MBPT), is essential. This study aimed to develop and evaluate machine learning (ML)-based predictive models incorporating multi-dimensional datasets to enhance diagnostic accuracy for asthma.

Data were analyzed from 1,501 individuals referred for suspected bronchial asthma to a tertiary care center from 2015 to 2021. Diagnoses were made by specialists based on comprehensive clinical evaluations. The dataset was split into training (75%, n=1,125) and test (25%, n=376) sets. A Gradient Boosting Machine (GBM) was used to create predictive models with features including demographics, symptoms, blood tests, FeNO, PFT, sputum analysis, and MBPT. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), and DeLong's test with Bonferroni post hoc adjustments was used for pairwise comparisons.

The model's performance progressively improved with the addition of more features. The initial model, which included demographic information and symptoms (Model 1), achieved an AUROC of 0.645 (95% CI: 0.586–0.704). When blood tests were added (Model 2), the AUROC increased to 0.735 (95% CI: 0.683–0.788). Further inclusion of FeNO levels (Model 3) raised the AUROC to 0.827 (95% CI: 0.782–0.871). Incorporating PFT results (Model 4) resulted in an AUROC of 0.890 (95% CI: 0.855–0.925), and adding sputum analysis results (Model 5) improved it to 0.903 (95% CI: 0.870–0.935). Including PC20 (Model 6) led to an AUROC of 0.917 (95% CI: 0.883–0.950). Finally, the most comprehensive model, which included MBPT results (Model 7), achieved the highest AUROC of 0.939 (95% CI: 0.913–0.965). It is noteworthy that the MBPT-only model also demonstrated strong performance with an AUROC of 0.899 (95% CI: 0.867–0.932), showing significant differences with Models 1, 2, and 7 (p < 0.001).

Incorporating multi-dimensional datasets into ML models improved asthma diagnostics, with Model 7 achieving the highest accuracy. These results highlight the importance of MBPT and other objective tests in asthma diagnosis.