Unrecognized atrial fibrillation (AFib) is a significant issue, one that grows more and more common as time goes on. A team of researchers worked to see if AI and patient electronic health record (EHR) data could play a role in diagnosing patients with AFib before its too late, sharing its findings in JAMA Network Open.
“A major challenge in the management of patients with AFib is that stroke is often the first presentation of AF, indicating that simply waiting for a patient to develop AFib may not be the optimal approach to limiting the risk of stroke,” wrote Premanand Tiwari, MS, University of Colorado School of Medicine in Aurora, and colleagues. “On the other hand, population-wide screening for AFib is not currently recommended, although some suggest that targeted screening may be useful. A model that could predict risk of six-month incident AFib could be applied to target screening and identify a patient with AFib before their next clinic visit.”
Tiwari et al. applied numerous machine learning-based prediction algorithms to patient EHR data in hopes of helping healthcare providers screen patients at a higher risk of AFib before a stroke or other serious complications occurred. To develop its AI models, the team explored data from 2.2 million patients from three Colorado facilities that share a single EHR. The patients all received care from Jan. 1, 2011, to Oct. 1, 2018. While 80% of the patient records were used for training purposes, the other 20% was used as a testing dataset.
All AI models were then compared with an “unregularized logistic regression model based on the presence of known clinical predictors of AFib,” including obesity, diabetes, coronary artery disease, heart failure and more.
“Among the approaches examined, we found that a single-layer shallow neural network using the 200 most common EHR features, including age and sex, was superior to other methods, which included regularized regression, gradient boosted descent, random forest and a deep neural network,” the authors explained.
This superior model achieved an area under the ROC curve of 0.800, a specificity of 84.9%, sensitivity of 75.2%, negative predictive value of 99.6% and positive predictive value of 5.9% “with relatively poor calibration across predicted probabilities.” This was only a slight improvement than the team’s logistic regression model based on clinical predictors of AFib, which had an area under the ROC curve of 0.794.
The machine learning model using EHR data may have not been “substantially better than a simpler model,” the authors noted, but it did represent “the optimal classification of risk.”
“We studied the development of an machine learning model for predicting six-month risk of AFib using harmonized EHR data and found that the combination of random oversampling and single-layer neural network classification provided superior prediction than other machine learning models,” the team concluded. “Further work is needed to explore the technical and clinical applications of this model to improving outcomes.”