Development and evaluation of a multivariate prediction model for diagnosing asthma in patients with clinically suspected asthma using capnography
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Diagnosis of asthma in primary care is challenged by a multistep pathway with variable adherence leading to significant misdiagnosis, late diagnosis and poor patient outcomes. This is driven by a lack of rapid, accurate, easy-to-use diagnostic tests that can reliably diagnose asthma with both high sensitivity and specificity. The objective of this work was to develop and evaluate a multivariate machine learning classifier for asthma diagnosis. The classifier was built using interpretable data processing and machine learning techniques applied to 75-second tidal breathing CO 2 recordings captured on TidalSense’s N-Tidal ® hand-held capnometer. The target population comprises patients with clinically suspected asthma who have no contraindications to performing capnometry.
Methods
Capnograms were collected from 138 asthmatic and 132 non-asthmatic participants (including healthy volunteers, and those with chronic obstructive pulmonary disease (COPD), heart failure, and other cardiores-piratory conditions) recruited from both primary and secondary care. Each high-resolution CO 2 recording was transformed into 82 features (using the N-Tidal ® Diagnose 1 v1.0 software) that characterise the constituent breathing cycles. A logistic regression model was trained on these features and performance metrics generated from an unseen test set of 64 participants. Model performance was evaluated using discrimination, measured by the area under the receiver operating characteristic curve (AUROC), as well as clinically relevant predictive accuracy metrics, including positive predictive value (PPV) and negative predictive value (NPV). This was repeated 20 times with different training and testing participants for additional statistical robustness; the average and variability of these metrics were recorded.
Results
The classification model achieved an AUROC of 0.91 ± 0.03%, sensitivity of 83 ± 4%, specificity of 85 ± 6%, positive predictive value (PPV) of 87 ± 4%, and negative predictive value (NPV) of 81 ± 4% in detecting asthma from a single breath recording. The model demonstrated diagnostic stability, with 95.8% of each participant’s recordings over the course of data collection being classified correctly on average. No model bias was observed with regards to sex, but performance did improve with age, possibly reflecting increasing severity of disease with age.
Conclusion
This study introduces a highly accurate and interpretable multivariate diagnostic model capable of classifying asthma from a single breath recorded using the N-Tidal ® Handset. It achieves high sensitivity and specificity compared with current methods, such as spirometry, and could enable point-of-care diagnosis in patients suspected of having asthma.
