Machine Learning Analysis of Microbiome Reveals Children’s Asthma Control Level

Machine learning gastrointestinal microbiome analysis of saliva and feces samples can be used to reveal controlled vs uncontrolled asthma in children with moderate-to-severe asthma, according to study findings published in Pediatric Allergy and Immunology.

Although research demonstrates that the microbiome may be associated with asthma characteristics, the association remains to be clarified. Researchers sought to determine whether the gastrointestinal microbiome can be used to discriminate between controlled and uncontrolled asthma in children. Secondarily, the researchers also sought to (1) discover a set of potential biomarkers for predicting asthma treatment responsiveness in children; and (2) compare the benefits of using machine learning vs conventional analysis techniques.

In the current study, researchers analyzed salivary and fecal microbiome data of the SysPharmPediA study (ClinicalTrials.gov Identifier: NCT04865575), a multicenter, prospective, European study conducted in 4 centers in Germany, Spain, Slovenia, and the Netherlands. That trial involved 143 children, from 6 to 17 years of age, diagnosed with controlled and uncontrolled asthma and being treated according to the Global Initiative for Asthma guidelines for step 3 or greater. These children all provided saliva and/or feces samples. Children in the trial with uncontrolled asthma at least 1 severe exacerbation within the 12 months prior to the study and a childhood asthma control test (cACT) score of at least 19; children with controlled asthma had no severe exacerbations and a cACT score of less than 19. Participants had a median age of 12 years, were predominantly boys (59.4%; uncontrolled 57.3%, controlled 63.0%) and were predominantly European.

Investigators for the current study took a multi-analytical approach to microbiome analysis that involved: (1) conventional global diversity measures (2) differential abundance analysis; and (3) advanced machine learning via recursive ensemble feature selection (REFS). Notably, the investigators did not find significant difference between children with controlled or uncontrolled asthma using global diversity and DAA methods. Using REFS, however, the researchers were able to detect Haemophilus and Veillonella taxa with an average classification accuracy of 81% (saliva) and 86% (feces) to differentiate controlled and uncontrolled asthma.

For both saliva and feces, these taxa revealed enrichment previously associated with inflammatory diseases that was also found in saliva samples with chronic obstructive pulmonary disease. Among children who had taken antibiotics in the previous 2 months, the analysis found 450 amplicon sequence variants (ASVs) detected in 143 patient saliva samples and 463 ASVs in 103 patient feces samples; in children who had not taken antibiotics, the analysis found 436 ASVs in 124 saliva samples and 451 ASVs in 87 feces samples.

Study limitations include: (1) the possibility that children in the controlled asthma group with an unknown risk of exacerbation could have had similar microbiome composition to children in the uncontrolled group; and (2) a sample size that was underpowered for machine learning techniques.

Significantly, said the researchers, “We showed that diversity measures and conventional differential abundance analyses were unable to discriminate between controlled and uncontrolled asthmatics, while the machine learning technique REFS was able to find sets of taxa with predictive power for asthma control status.” The study authors concluded that “The gastrointestinal microbiome can be used to discriminate between controlled and uncontrolled asthmatics using machine learning.”

Disclosure: Some study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures.