Predicting Treatment Outcomes Using Explainable Machine Learning in Children with Asthma

Area of Science:

• Pediatric Pulmonology
• Computational Biology
• Immunology

Background:

• Childhood asthma is a complex, heterogeneous disease with variable treatment responses.
• Current anti-inflammatory treatments are inadequate for a significant number of pediatric asthma patients.
• Understanding predictors of treatment success is crucial for optimizing management strategies.

Purpose of the Study:

• To employ machine learning algorithms for predicting treatment success in pediatric asthma.
• To identify key variables influencing treatment outcomes for improved mechanistic understanding.
• To evaluate prediction accuracy based on asthma control, lung function (FEV1, MEF50), and fractional exhaled nitric oxide (FENO).

Main Methods:

• Utilized Random Forest and AdaBoost classifiers on a cohort of 365 children with mild to severe asthma.
• Assessed treatment outcomes after 6 months of controller medication based on changes in asthma control, FEV1, MEF50, and FENO.
• Identified predictive variables associated with treatment response.

Main Results:

• Machine learning models achieved higher prediction power for asthma control and FENO-based outcomes, especially in younger children.
• Asthma severity and total IgE were significant predictors for both asthma control and FENO outcomes.
• MEF50-based outcomes were better predicted than FEV1-based responses, with hsCRP highlighting distal airway involvement.

Conclusions:

• Machine learning offers a robust tool for predicting treatment success in pediatric asthma, outperforming traditional lung function metrics.
• Asthma control and FENO-guided management, complemented by machine learning predictions, support precision medicine in pediatric asthma.
• T2-high asthma phenotype appears to respond best to anti-inflammatory therapies, warranting further investigation.