Data-efficient machine learning approach for predicting asthma attack risk

Asthma caused around 436,000 deaths globally in 2021. Predicting asthma attacks can save lives, reduce costs, and strengthen healthcare systems. However, risk prediction requires extensive data, which is often unavailable. This study aims to develop an approach that minimizes data requirements, reducing complexity and enabling broader applicability in data-limited settings. We employed the CRISP-DM methodology, including combinations of feature selection strategies, machine learning algorithms, and data imbalance handling techniques. Altogether, 120 models were constructed. Model performance was evaluated based on accuracy and efficiency. The top models were externally validated. The key finding is that the LR and XGB models, when combined with an under-sampling technique, perform better, given that the feature selection is conducted based on the feature importance generated by the XGB-RUS model. Even a few features, including asthma attacks in the past year and SABA_ICS ratio, can achieve a reasonable level of performance. Through rigorous experimentation, we achieved high accuracy in multiple scenarios. We have presented 6 feature sets. Among those, FS1 achieved the best performance, while FS6 yielded a reasonable performance with only two features. This paper presents a comprehensive analysis of the data requirements for predicting the risk of asthma attacks. This will enhance the acceptability of the prediction models using simplified feature sets in clinical settings.