Unsupervised identification of asthma symptom subtypes supports treatable traits approach

Heterogeneity of asthma requires a personalized therapeutic approach. However, objective measurements, such as spirometry and fraction of exhaled nitric oxide (FeNO) for implementing treatable traits approach, are limited in low- and middle-income countries and non-specialist settings. To implement precision medicine even with minimal resources, we developed an algorithm using unsupervised machine learning techniques that estimates key treatable traits (airflow limitation, type 2 [T2] inflammation, and frequent exacerbations) based on an asthma patient-reported outcome (PRO).

We applied hierarchical clustering and Uniform Manifold Approximation and Projection (UMAP) to Asthma Control Questionnaire (ACQ)-5 including five residual symptoms from two asthma cohorts (the discovery cohort with 1697 patients and validation cohort with 157 patients).

We identified five symptom clusters, characterized by key treatable traits: Cluster 1, minimal asthma symptoms; Cluster 2, a little symptom, mild airflow limitation; Cluster 3, predominant shortness of breath and wheezes, airflow limitation; Cluster 4, predominant morning symptoms and nocturnal awakening, T2 inflammation; and Cluster 5, all symptoms severe, airflow limitation, T2 inflammation and frequent exacerbations. The UMAP projections of ACQ-5 (five-dimensional) to two-dimensions allowed to visualize datapoints and clusters, which visually revealed that patients with poorly-controlled asthma were divided into Clusters 3, 4 and 5. These results were externally validated in an independent cohort.

Based on asthma PRO data, the developed algorithm categorized asthma patients into five symptom-based subtypes that provide insights into key treatable traits. Our data-driven digital health approach will extend precision medicine of asthma to medical facilities even in resource-constrained settings.