Application of ML methods in identifying patients with asthma in primary care

Abstract

Abstract Background: Asthma is one of the most prevalent diseases, with approximately 5.4 million patients on prescribed medication in the UK. Poor asthma management is responsible for many preventable deaths in the UK, making the mortality rate the highest in Europe. Identifying asthma patients is time-consuming and requires detailed reviews of individual patients by GPs. In a previous study (awaiting publication), bespoke designed algorithms (Smart-Searches™) were used to identify patients who were not on the Quality Outcome Framework (QOF) asthma register but were likely to have asthma. GPs further reviewed these patients found by the searches to confirm their condition. This study aims to apply machine learning methods to real-world primary care electronic health records (EHRs) and compare their performance in identifying asthma patients with the previously used Smart-Searches™. Methods: This is a binary classification problem where patients are identified as asthmatic or non-asthmatic. Data from two practices used in this study comprised around 9000 patients, of whom around 600 were on the asthma register. A set of 40–45 features were extracted from the health records as inputs to the models. The models were trained and tested on datasets in several experiments. Both linear models such as Logistic Regression, Random Forest, Support Vector Model, Naïve Bayes, and deep learning models such as MLP and CNN were evaluated, and compared with the existing traditional methods. Results: ML models, on average, got a higher accuracy of about 70% compared to traditional methods at 54%. The Ensemble model obtained the highest accuracy at 77%, followed by MLP at 75%. In addition, the average positive predictive value for the ML methods was 82% compared to the search-based system at 54%. Finally, the Naïve Bayes model obtained the best positive predictive value at 100%. Conclusions: ML methods obtained high accuracy and positive predictive values, showing that the ML models could make better asthma identification predictions than the existing system. This also shows that the machine learning models could help clinicians identify more asthma patients in significantly less time while requiring less clinician input than the existing best methods leading to improved efficiency and better patient care.