Using clusters of asthmatic patients to quantitatively estimate, prior to initiating treatment, the likelihood of β2-agonist response in asthma patients

Abstract

Background: Estimating the probability of an asthma patient’s response to long-acting beta-2 adrenergic receptor agonists (LABA) from only a few lung function measurements would help clinicians avoid unnecessary drug prescription, contributing to personalized treatment and reduced costs.

Aims and objectives: To develop a computational methodology capable of generating estimates of the probability of LABA-responsiveness and assess its accuracy using cohort data.

Methods: Using fluctuation-based clustering (Delgado-Eckert E. et al. Thorax 2018; 73.2:107-115) we retrospectively analysed time series of PEF measurements recorded by 79 mild-to-moderate asthmatic adults during the placebo phase of a study aimed at characterizing the effects of salmeterol and salbutamol treatment (described in Thamrin C. et al. Eur Respir J 2009; 33:486–493). This yielded three clusters of patients: 1, composed mainly of salmeterol responders; 2, mainly non-responders; 3, a mixed group. We randomly removed 99% of the data from each time series, using these stripped-down versions as proxies for new patients’ data. Using Softmax Regression, we calculated the probability that a patient would be assigned to one of the three clusters. Based on the proportion of responders in each cluster, we calculated the salmeterol-responsiveness probability of a given “new” patient. We used a Receiver Operating Characteristic (ROC) curve to assess the probability’s classification performance.

Results: The LABA-responsiveness probability ROC curve had a mean area under the curve of 0.7.

Conclusion: Our method yields LABA-responsiveness probability estimates with good discriminatory power.