Key result
An automated breath classification system using a Decision Tree classifier achieved an AUC of 90% for detecting abnormal capnogram breaths, enabling robust pre-processing of respiratory waveforms.
Effect estimate: AUC 90%
An automated system using simple classifiers can accurately segment and classify capnogram breaths, providing robust pre-processing for reliable respiratory variability analysis.
May enhance respiratory waveform pre-processing in studies; leaves open clinical validation before monitoring use.
Variability analysis of respiratory waveforms has been shown to provide key insights into respiratory physiology and has been used successfully to predict clinical outcomes. The current standard for quality assessment of the capnogram signal relies on a visual analysis performed by an expert in order to identify waveform artifacts. Automated processing of capnograms is desirable in order to extract clinically useful features over extended periods of time in a patient monitoring environment. However, the proper interpretation of capnogram derived features depends upon the quality of the underlying waveform. In addition, the comparison of capnogram datasets across studies requires a more practical approach than a visual analysis and selection of high-quality breath data. This paper describes a system that automatically extracts breath-by-breath features from capnograms and estimates the quality of individual breaths derived from them. Segmented capnogram breaths were presented to expert annotators, who labeled the individual physiological breaths into normal and multiple abnormal breath types. All abnormal breath types were aggregated into the abnormal class for the purpose of this manuscript, with respiratory variability analysis as the end-application. A database of 11,526 breaths from over 300 patients was created, comprising around 35% abnormal breaths. Several simple classifiers were trained through a stratified repeated ten-fold cross-validation and tested on an unseen portion of the labeled breath database, using a subset of 15 features derived from each breath curve. Decision Tree, K-Nearest Neighbors (KNN) and Naive Bayes classifiers were close in terms of performance (AUC of 90%, 89% and 88% respectively), while using 7, 4 and 5 breath features, respectively. When compared to airflow derived timings, the 95% confidence interval on the mean difference in interbreath intervals was ± 0.18 s. This breath classification system provides a fast and robust pre-processing of continuous respiratory waveforms, thereby ensuring reliable variability analysis of breath-by-breath parameter time series.
No takes yet. Share an insight, caveat, or question.
Herry et al. (2014) studied Respiratory monitoring (n=300). Automated breath classification system vs. Expert visual analysis was evaluated on Classification performance (AUC) for detecting abnormal breaths (AUC 90%). An automated breath classification system using a Decision Tree classifier achieved an AUC of 90% for detecting abnormal capnogram breaths, enabling robust pre-processing of respiratory waveforms.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: