Leveraging Machine Learning to Enhance Asthma Phenotyping: Implications for Precision Medicine

Written by Chao-Ping Wu, MD

Asthma is a multifaceted respiratory condition characterized by chronic airway inflammation and variable respiratory symptoms. It’s diagnosis and management can be particularly challenging due to its heterogeneous nature. Historically, asthma phenotyping has been attempted using well-organized clinical registries. However, concerns persist regarding the applicability of these findings to real-world settings, as clinical registries often involve selective patient groups that may not represent the broader asthma population.

In contrast, our study, which appeared earlier this year in The Journal of Allergy and Clinical Immunology, leveraged machine learning (ML) and extensive real-world data to explore asthma's complex heterogeneity more comprehensively. This approach allows for the characterization of asthma at a granularity that reflects true clinical scenarios, thus enhancing the relevance and applicability of the phenotypes identified in everyday practice.

The necessity for advanced asthma phenotyping

Traditional asthma classification systems, which primarily assess severity based on symptom frequency and intensity, often do not correspond with the actual burden of disease experienced by patients. For instance, individuals classified with mild asthma might frequently require acute care, illustrating a significant discrepancy between clinical classifications and their real-world implications.

This misalignment highlights the urgent need for sophisticated phenotyping methods that capture the full spectrum of clinical manifestations and underlying biological mechanisms of asthma. Past research has identified multiple asthma subphenotypes, including eosinophilic, paucigranulocytic and neutrophilic types. Each of these is associated with distinct inflammatory pathways and clinical outcomes. Recognizing and differentiating these subphenotypes is crucial not only for assessing disease severity more accurately but also for predicting therapeutic responses and guiding personalized treatment strategies, potentially including early intervention with biological therapies to halt disease progression.

Study findings and phenotype characteristics

Our study employed an innovative unsupervised machine learning approach to analyze electronic health records (EHRs) from 13,498 asthma patients at Cleveland Clinic, focusing on a range of clinical and demographic variables. We utilized k-prototype clustering to identify five distinct asthma subphenotypes, each with unique characteristics that challenge traditional correlations between clinical severity and exacerbation frequency. These subphenotypes also suggest underlying inflammatory processes not detected by standard assessments. They include:

1. Young, Non-Severe Asthma: Predominantly younger individuals with frequent exacerbations despite normal lung function tests. This phenotype challenges the traditional correlation between clinical severity and exacerbation frequency, suggesting underlying inflammatory processes not detected by standard assessments.
2. Obese Female: This phenotype predominantly comprises Black women with obesity who experience high rates of both severe and non-severe asthma exacerbations. These are intensified by obesity-induced inflammation and hormonal effects on IL-17 production. These patients frequently require high-dose inhaled corticosteroid (ICS) therapy and often face challenges related to depression. The inclusion of pregnant women in this cluster underscores the importance of specialized asthma management during pregnancy to address the unique risks and complexities associated with their condition.
3. Elderly, Non-Severe Asthma: Consisting mostly of older adults with low eosinophil counts and maintained normal lung function, this cluster is at a low risk for severe exacerbations. The findings suggest a possible link between menopause and reduced asthma severity, highlighting the need for further research into hormonal changes during menopause and their impact on asthma.
4. Fixed Airway Obstruction: This group is comprised of patients with lung function indicative of fixed airway obstruction that is likely due to chronic inflammation and airway remodeling.
5. Severe, Eosinophilic Asthma: Characterized by high eosinophil counts and severe symptoms, patients in this cluster often exhibit steroid resistance, necessitating alternative treatments such as biologics. This phenotype is particularly challenging to manage due to its resistance to conventional therapies and frequent exacerbations.

Impact on knowledge and clinical practice

The identification of these subphenotypes significantly enriches our understanding of asthma's clinical spectrum and facilitates the implementation of phenotype-specific management strategies. By aligning treatment approaches with phenotypic characteristics, clinicians can achieve better therapeutic outcomes, reduce unnecessary healthcare utilization and advance personalized medicine in asthma care. These findings have the potential to transform the standard of care at Cleveland Clinic and similar institutions by providing a more precise, patient-tailored approach to asthma management.

Conclusion

This study underscores the potential of machine learning to revolutionize asthma phenotyping using real-world data to develop more precise and effective management across its diverse spectrum. By shifting towards a data-driven, personalized approach, we can anticipate significant improvements in treatment regimens and quality of life for patients with asthma. Our findings advocate for ongoing research to further refine these phenotypes and explore new therapeutic targets that address the specific needs of each identified patient group.