Introduction

The aim of personalized medicine is to provide a target therapy for each individual or phenotype, based on the corresponding syndrome or disease1. Even though machine learning techniques have identified a number of structures and/or phenotypes in asthma, one has to be careful in clinical interpretation of these structures, as they may not represent true endotypes (underlying immunopathological mechanisms)2. Overlaps in endotypes as well as clinical presentation of the disease make the delivery of personalized asthma treatment quite elusive3,4. Furthermore, the same pattern of symptoms does not necessarily indicate the same underlying mechanism and moreover, different mechanisms are not mutually exclusive and may even act synergistically.
The emergence of advanced machine learning algorithms, abundance of clinically significant data and computing power can be attributed as key enablers in the development of personalized medicine. There is a substantial body of scientific work on data-driven methods in asthma phenotyping and the variables as well as the model chosen can largely affect the models and obtained results5-11. Careful selection of both the predictive model and the dataset are essential in such studies, with expert clinical interpretation being of utmost importance.
In childhood asthma, inhaled corticosteroids- ICS (in combination with long-acting beta-agonists, LABA and/or add-on leukotriene receptor antagonists, LTRA) remain controller medications of choice, although evidence shows that a significant proportion of patients fail to respond adequately to such treatment.5–8The complexity of the disease, or better said, the “umbrella” diagnosis of asthma that encompasses a number of different phenotypes underpinned by different pathophysiological mechanisms or distinct inflammatory pathways (endotypes) seems to be the major obstacle in asthma management as well as in the development of personalized treatment approaches9.
An important study on prediction was conducted by Belgrave et al.10, focusing on preschool wheezers (N=150) with a large dataset (N(variables)=636) using selected state-of-the-art techniques for data processing and machine learning and obtained 90%+ performance in Kappa statistics. The authors also reported robustness and performance quality when using Random Forest and that subjective variables are important in distinguishing ill patients from controls.
A key study focusing on asthma control-based response to controller medication was conducted in the Childhood Asthma Management Program (CAMP) cohort using novel machine learning algorithms11. They reported that asthma control, bronchodilator response and serum eosinophils were the most predictive features of asthma control, regardless of the medication used.

Our aim and contribution

In this work we present the data and results from an observational childhood asthma study in an ethnically homogeneous, age diverse cohort, reflecting real-life clinical situations with the majority of patients having mild to moderate disease. The primary aim was to test the predictive possibilities for treatment success in pediatric asthma patients and reveal key variables for understanding the mechanisms underlying such response. The primary endpoints were four different parameters of response to treatment, assessed by changes in lung function i) Forced expiratory volume in one second (FEV1) and ii) Maximal Expiratory Flow at 50% of Vital Flow Capacity (MEF50), iii) changes in airway inflammation- Fractional Exhaled nitric oxide (FENO) and iv) the level of asthma control assessed by a pediatric pulmonologist. Each of these targets was evaluated at baseline and after 6 months of treatment use alongside other parameters and biomarkers. The predictive possibilities were tested using the Random Forest (RF) and Adaptive Boosting (AdaBoost) machine learning algorithms. Since these algorithms are considered to be black-boxes, we introduced model explainability by the use of variable importance for evaluating the most important variables for differentiating between responders and non-responders. Furthermore, we discuss the use of different classification metrics for the validation of the results.