Machine Learning Can Play a Role in Predicting Acute Asthma Exacerbation, Study Finds

It can be difficult for clinicians and patients to predict when an acute asthma exacerbation might occur. But findings from one new study suggest that providers may be one step closer to predicting patients’ risk of experiencing an asthma episode simply by viewing their electronic health record (EHR).

In a new study, Cleveland Clinic researchers demonstrated that machine learning (ML) can predict the occurrence of a non-severe asthma exacerbation, an asthma-related ED visit or hospitalization. They published their findings in CHEST.

“In application,” says Joe Zein, MD, PhD, a pulmonologist in the Respiratory Institute and the study’s first author, “this means the ML algorithm could interface with patients’ EHR, interpret their unique demographic information, clinical data and biomarkers and make a prediction about the probability of an asthma episode within a certain time frame.”

Previous studies using traditional statistical modeling and known disease risk factors performed sub-optimally, leading Dr. Zein and his colleagues to explore the feasibility of ML in this context. While ML is increasingly utilized in medicine, even in pulmonary medicine, no other studies to date have examined the risk of an asthma episode using ML and ambulatory data extracted from patients’ EHR.

Study design and primary findings

In the retrospective review, investigators extracted data from a cohort of Cleveland Clinic patients (N = 60,302), followed from 2010 to 2018. The patients’ ages ranged from 18 to 80, and all of them reported using asthma rescue or controller medication for more than six months. The study excluded pregnant women, active smokers or those with a significant smoking history, and patients with any other chronic pulmonary diseases.

Using real-life, readily available data on demographic and clinical information extracted from EHRs, they developed three predictive models — one using traditional statistical modeling and two using an ML algorithm: Light Gradient Boosting Machine (LightGBM) and random forest regression method. In every testing scenario, says Dr. Zein, the ML models often performed better than the traditional model, but LightGBM performed the best.

Performance results of LightGBM show an area under the receiver operator curve of 0.71 (95% CI: 0.70-0.72), 0.88 (95% CI 0.86-0.89) and 0.85 (95%CI: 0.82-0.88), respectively, when predicting the use of oral glucocorticoid burst, an asthma-related ED visit and hospitalization. The investigators were also able to replicate these findings in a subset validation study with an independent database of patients, followed from 2019 to 2020.

Machine learning: A tool in clinical management

The findings of this study were not surprising to Dr. Zein, who asserts that ML has great potential in medicine, particularly in the delivery of highly personalized care. “We know that ML handles missing values and large data sets better than classic statistical modeling. It’s often more predictive because of its capacity to leverage real-world data.”

He continues, “In essence, ML algorithms are designed to make more accurate predictions models than classic statistics. It uses learning algorithms to find patterns in often large and cumbersome data. It makes minimal assumptions about the methods of data collection and is effective even when the data are collected without stringent experimental designs. Equally important is the fact that ML algorithms account for complex nonlinear interactions between data variables.”

In this case, Dr. Zein says, the long-term goal is to use ML to interface with an EHR to help inform their clinical management decision tools. The next logical step in bringing this application to fruition is conducting a prospective study that evaluates the performance of ML algorithms. He also affirms the importance of context and human interpretation.

“We are pleased that our findings demonstrate that ML can be used to build decision-making tools in the clinical management of asthmatic patients,” he says. “ML is effective in aiding with the integration of big data, but the physician provides judgment rooted in clinical context and experience,” he says. “Both are extremely important.”