Discovering Pediatric Asthma Phenotypes on the Basis of Response to Controller Medication Using Machine Learning

Area of Science:

• Pediatric pulmonology
• Computational biology
• Clinical informatics

Background:

• Pediatric asthma exhibits significant heterogeneity in inflammation and symptom control, necessitating personalized treatment strategies.
• Existing clustering methods for identifying asthma phenotypes and predicting outcomes have limitations, particularly in analyzing treatment response.
• Machine learning (ML) approaches for long-term pediatric asthma outcome prediction require further investigation.

Purpose of the Study:

• To utilize a novel ML algorithm, predictor pursuit (PP), to discover pediatric asthma phenotypes based on controller medication response.
• To predict longitudinal asthma control in children with asthma using the PP algorithm.
• To identify key clinical features associated with asthma control within each identified phenotype.

Main Methods:

• Applied the predictor pursuit (PP) ML algorithm to the Childhood Asthma Management Program (CAMP) dataset (n=1,019).
• Discovered phenotypes based on asthma control differences between budesonide and nedocromil treatment groups.
• Validated PP's phenotype discovery capability using the Asthma Clinical Research Network/Childhood Asthma Research and Education (ACRN/CARE) network data and compared prediction performance against traditional ML methods.

Main Results:

• Identified four consistent pediatric asthma phenotypes: allergic/non-obese (A+/O-), obese/non-allergic (A-/O+), allergic/obese (A+/O+), and non-allergic/non-obese (A-/O-).
• Observed differential treatment responses, with non-obese children responding better to budesonide and obese children to nedocromil.
• Demonstrated that the PP algorithm significantly outperformed traditional ML algorithms (P < 0.001) in predicting both short- and long-term asthma control.

Conclusions:

• The predictor pursuit (PP) algorithm effectively identifies distinct pediatric asthma phenotypes based on treatment response.
• These phenotypes offer a framework for personalized asthma management and improved long-term outcome prediction.
• Advanced ML methods, like PP, are valuable tools for dissecting complex diseases such as asthma and guiding treatment decisions.