Written in collaboration with Nyx Robey

Recent developments in machine learning for asthma prediction

As an asthma forecasting company, we strive to reduce the burden of asthma so that patients can better prepare for the possibility of an exacerbation. We do so based on forecasting asthma control based on weather, pollution, previous spirometry measures, and most recently, previous days of asthma control.

What is Asthma Control?

Asthma control as defined by the National Lung, Health, and Blood Institute involves (1) reducing impairment, or the frequency and intensity of symptoms and limitations, and (2) reducing risk, or the likelihood of future asthma attacks or worsening of lung function.[1] It is frequently used by clinicians to determine the degree of asthma care required by patients and how much such care should be periodically adjusted. It is, in other words, a prime indicator of how well a patient’s lungs are doing and is a key framework for making clinical decisions.

What does the Research Say?

In recently published respiratory literature, machine learning has previously been used to predict whether a patient will be diagnosed with asthma[2], or help triage patients who are in an emergency room[3]. However, recently, two studies have been published that predict an asthma patient’s functioning anywhere from a day to a week in advance, allowing for the possibility of live remote monitoring of patients by their physicians.

One, an international study,[4] predicted whether a patient would experience an asthma exacerbation within the next day, utilizing patient PEF and symptom scores. The second study predicted whether patients would exhibit high or low levels of asthma symptomatology[5]. The last study predicted asthma control deterioration in children one week in advance[6].

What if there’s no spirometer?

All of these models take spirometry measurements as predictors; however, we wanted to create a novel model that makes achieving high asthma control possible even for patients who cannot access a spirometer. We’ve started this process by validating our asthma control survey against industry standards — and now we’re also applying it through creating models that can forecast control a day in advance without requiring a spirometer.

Asthma control can be tracked and predicted through recording asthma symptoms

At VitalFlo, we recently created a machine learning model that predicts a patient’s level of daily asthma control around a day in advance without requiring spirometry measurements. This model performs almost identically well to our model that includes spirometry measurements, demonstrating the primacy of past symptoms in helping predict future control. Using symptoms alone may prove just as effective as using more granular measures such as spirometry.

The model allows us to see if patients’ symptomatology will change from day to day. These criteria were created using NIH asthma control guidelines[2], and data were collected from adolescent asthmatic patients from a symptom survey sent out via the VitalFlo smartphone app.

Preliminary Results from Our Model

Three previously published models for predicting asthma exacerbation or control reported accuracy ranging from 71–100%*, specificity ranging from 71–100%**, and recall ranging from 73–100%***.

Our model performs comparably: we’ve achieved an accuracy of 93% with accompanying specificity and recall of 97% and 72%, respectively. This performance is identical to our model that uses spirometry measurements.

What does this mean for the future of remote lung health monitoring?

The results are promising. Applying and deploying this machine learning model may help physicians successfully monitor lung health for a broader range of respiratory patients — those who cannot use spirometers, or do not have access to them. Understanding lung function triggers and symptoms will ideally help improve patients’ control of their asthma. This preliminary exploration suggests that providing demographics, collecting information about air quality, and completing symptom surveys may provide such control improvement.

We’re excited to continue to draw insights from machine learning that might support physicians in making the best possible patient care decisions.

Footnotes

*Accuracy: how many of the model’s predictions are correct?

** Specificity: of all the actual days without symptoms, which did the model predict were days without symptoms?

*** Recall: of all the actual days with symptoms, which did the model predict were days with symptoms?

References

1. National Lung, Health, and Blood Institute. Asthma care quick reference: Diagnosing and managing asthma.
2. Spathis, D., & Vlamos, P. (2019). Diagnosing asthma and chronic obstructive pulmonary disease with machine learning. Health Informatics Journal, 25(3), 811–827.
3. Patel, S. J., Chamberlain, D. B., & Chamberlain, J. M. (2018). A machine learning approach to predicting need for hospitalization for pediatric asthma exacerbation at the time of emergency department triage. Academic Emergency Medicine, 25(12), 1463–1470.
4. Zhang, O., Minku, L. L., & Gonem, S. (2020). Detecting asthma exacerbations using daily home monitoring and machine learning. Journal of Asthma, 1–10.
5. Finkelstein, J., & Jeong, I. (2017). Machine learning approaches to personalize early prediction of asthma exacerbations. Annals of the New York Academy of Sciences, 1387(1), 153.
6. Luo, G., Stone, B. L., Fassl, B., Maloney, C. G., Gesteland, P. H., Yerram, S. R., & Nkoy, F. L. (2015). Predicting asthma control deterioration in children. BMC Medical Informatics and Decision Making, 15(1), 1–8.

‍