Machine learning model predicts small airways disease in patients with asthma
Click Here to Manage Email Alerts
Key takeaways:
- A machine learning classifier was trained on capnogram features.
- The model’s probability output inversely correlated with percent-predicted forced expiratory flow rate between 25% and 75% of vital capacity.
SAN DIEGO — A model that applied machine learning to capnography signals accurately detected small airways disease among patients with asthma, according to a presentation at the American Thoracic Society International Conference.
Small airways disease contributes to asthma, tripling the odds for systemic corticosteroid use and increasing odds for acute exacerbations by a factor of six, Henry Broomfield, MSc, machine learning scientist, TidalSense, and colleagues wrote.
However, the researchers continued, the association between small airways disease and asthma is poorly understood. A low percent-predicted forced expiratory flow rate between 25% and 75% of vital capacity (FEF25-75) measured via spirometry indicates small airways obstruction.
“The objective of this research was therefore to use quick point-of-care CO2 monitoring using the N-Tidal (TidalSense) capnometer to assess small airways disease and ultimately improve treatment pathways for patients,” Broomfield said during his presentation.
The study included 85 patients with asthma in primary and secondary care in the United Kingdom participating in the longitudinal, observational and clinical Asthma Breathing Record Study and General Breathing Record Study.
Patients breathed normally into the handheld N-Tidal capnometer for 75 seconds. The capnometer records relaxed tidal CO2 concentrations. After the capnography signals were denoised, they were translated into 25 geometric features.
“This signal can be analyzed to make inferences about an individual’s cardiorespiratory health,” Broomfield said.
“The main use case thus far has been in COPD diagnosis,” he continued, “though there is plenty of information contained within the signal we would like to explore, and today that is more specifically small airways disease.”
Broomfield said that the accuracy of the sensor enables these inferences. The N-Tidal also is the first handheld capnometer that deals with condensation issues from direct sampling from users’ breath, he continued.
“We pick up a lower line that indicates greater accuracy, showing the N-Tidal is more accurate than current gold standards,” he said.
The N-Tidal also samples at a rate that is “vastly higher” than 10 kHz, Broomfield said.
“This allows it to detect fine-grained respiratory physiologic changes that have not been previously possible to see,” he said.
CO2 concentrations in the mouth increase during exhalations and return to background level during inhalations.
“We isolate each individual breath cycle, and then we split the breath into different phases,” Broomfield said. “We construct a set of features from this geometric waveform.”
Many of these features have been described in previous literature, Broomfield added, and they are related to underlying physiologic processes.
To quantify small airways obstruction, the researchers used a dataset of 36,446 capnograms from the 85 patients in the study.
“We can use these digital biomarkers to do a little bit of machine learning,” Broomfield said.
Using features from the capnograms of 82% of the patients, the researchers trained the XGBoost explainable machine learning classifier to distinguish percent-predicted FEF25-75 totals of less than 50% from those totals of 50% and higher.
“This is a commonly used biometric definition of small airways obstruction,” Broomfield said. “And this classifier was strong.”
The researchers then tested the classifier with 10 capnograms from each of the remaining 18% of patients.
The model had a 93% area under the receiver operating characteristic curve, 95% sensitivity, 76% specificity, 88.8% positive predictive value and 88.4% negative predictive value.
“The machine learning model actually gives us a prediction probability,” Broomfield said. “This is a probability of small airways disease.”
When the researchers plotted the average machine learning model prediction probability output per participant against the average percent-predicted FEF25-75 per patient, the Pearson’s product moment correlation coefficient between these variables was –0.902.
Overall, the researchers said, the machine model’s probability output inversely correlated with percent-predicted FEF25-75.
Broomfield cautioned that these findings are limited because they came from a cohort of patients who solely had asthma from a couple of regions with restricted demographics in the U.K.
“Ensuring generalizability amongst a more diverse group and in different disease cohorts will be the focus of future reports,” he said.
But based on these findings, the researchers concluded that the application of machine learning techniques to processed capnography signals could yield accurate indicators of small airways disease among patients with asthma.
“Thus, it could be used as an effort independent, accurate, rapid point-of-care test that identifies small airways disease and subtype asthma disparity,” Broomfield said.
The researchers also are investigating whether the system could be used to assess whether biologic therapies can eliminate mucus plugging in the very small airways of patients with asthma, he added, along with assessing correlations between the system’s findings and forced oscillometry and fractional exhaled nitric oxide.