Machine Learning Helps Identify Variables of Corticosteroid Response in Asthma

Researchers have found 12 variables that may be predictive of corticosteroid response in patients with severe asthma, according to a study published in the American Journal of Respiratory and Critical Care Medicine.

A total of 100 clinical, physiologic, inflammatory, and demographic variables from patients with asthma (N=346) who participated in the Severe Asthma Research Program underwent multiple kernel k-means clustering. These variables included paired sputum cell counts, obtained before and 2 to 3 weeks after receiving 40 mg triamcinolone. A 2-step feature selection machine learning procedure was used to select top predictive baseline variables that were predictive for cluster assignment.

There were 4 asthma clusters with varying responses to corticosteroids after multiple kernel clustering. The first 2 clusters included young, highly allergic asymptomatic patients with asthma with normal lung function and modest responses to corticosteroids. Clusters 1 (n=81) and 2 (n=73) were mostly separated by differing percentages of sputum neutrophil and macrophage percentages after receipt of steroids. Cluster 3 (n=96) included patients with late-onset asthma with low lung function, which also included patients with high baseline eosinophilia and high responsiveness to corticosteroids. In addition, cluster 4 (n=96) included young obese women with little eosinophilic inflammation, severe airflow limitation, and worse response to corticosteroid therapy.

Baseline variables identified as predictive of corticosteroid response included age, responses to activity limitation score on the asthma quality-of-life questionnaire, body mass index, baseline prebronchodilator forced vital capacity percent predicted, sputum macrophage percentages, number of specific immunoglobulin E antibodies, forced vital capacity albuterol response, baseline pulse, total white blood cell count, race, and sputum eosinophil percentages.

“The 12 predictive variables we identified suggest software can be developed which predicts responses to [corticosteroids], helping to make precision medicine possible,” the researchers wrote. “These machine learning approaches provide novel insight into [corticosteroid] response patterns that could improve asthma management.”

Reference

Wu W, Bang S, Bleecker ER, et al. Multiview cluster analysis identifies variable corticosteroid response phenotypes in severe asthma [published online January 25, 2019]. Am J Respir Crit Care Med. doi:10.1164/rccm.201808-1543OC