Is Artificial Intelligence Coming for Your Job?
By Michael H. Crawford, MD
SYNOPSIS: A retrospective analysis of plain chest X-ray images in the medical record using deep learning in patients suitable for risk assessment for atherosclerotic cardiovascular disease (ASCVD) has shown similar results as the American College of Cardiology/American Heart Association ASCVD risk calculator for determining who is at sufficient risk to consider statin therapy.
SOURCE: Weiss J, Raghu VK, Paruchuri K, et al. Deep learning to estimate cardiovascular risk from chest radiographs: A risk prediction study. Ann Intern Med 2024; Mar 26. doi:10.7326/M23-1898. [Online ahead of print].
Application of the American College of Cardiology (ACC)/American Heart Association (AHA) atherosclerotic cardiovascular disease (ASCVD) risk calculator often is hampered because some of the nine input variables are missing.
Thus, opportunistic automated risk assessment using commonly available tests in the electronic medical record, such as chest X-rays (CXRs), could be helpful.
Since deep learning (DL) has been shown to accurately estimate coronary artery calcium from CXR, these investigators from Germany and Boston sought to assess whether DL can predict the risk of major adverse CV events (MACE) from a routine CXR image.
The CVD risk model was developed using CXR data from a cancer screening study conducted between 1993-2001 from which a random subset of 40,718 subjects (80%) were used.
CV death was the endpoint. The derived CXR CVD risk model was externally validated in a separate group of outpatients from two hospitals in Boston who had a CXR taken between 2010 and 2011.
The model then was tested in patients aged 50-75 years who had a CXR, were not on a statin, were not diabetic, had a low-density lipoprotein (LDL) cholesterol between 70 mg/dL and 190 mg/dL, and had no history of MACE or ASCVD.
Finally, they assessed two cohorts: the first was 8,869 patients in whom one or more of the nine variables in the ASCVD risk score was not available (unknown ASCVD risk group) and the second was 2,132 patients with all variables available (known ASCVD risk group). These patients were grouped into categories using standard thresholds (low, < 5%; borderline, ≥ 5% to < 7.5%; intermediate, ≥ 7.5% to < 20%; and high ≥ 20%).
The primary outcome was MACE within 10 years of the CXR. The patients in both groups were largely white, the majority were women (55% and 65%, respectively), and their mean age was 60 years.
In the unknown ASCVD risk group, those with a CXR-predicted risk of ≥ 7.5% had a higher 10-year risk of MACE (adjusted hazard ratio [HR], 1.73; 95% confidence interval [CI], 1.47-2.03). In those with known ASCVD risk, the model predicted MACE beyond the traditional ASCVD risk score (HR, 1.88; 95% CI, 1.24-2.85).
If statin eligibility is defined as an ASCVD risk of ≥ 7.5%, the risk estimates from the CXR model and the traditional risk score were concordant for 70% of the patients and when discordant, the CXR risk score upclassified half the patients as statin-eligible.
In those with unknown risk, the CXR risk score resulted in a greater net benefit than treating everyone and treating no one. In those with a known risk score, the ASCVD risk score had a somewhat lower net benefit than the CXR score.
Cumulative MACE incidence curves for CXR CVD risk and the traditional ASCVD calculator were similar at the four levels of ASCVD risk.
The authors concluded that a single CXR may help identify individuals at high risk of MACE in whom the ASCVD risk calculator cannot be used because of missing data.
Commentary
The ACC/AHA ASCVD risk calculator was developed for adults aged 40-75 years who were not diabetic and had an LDL cholesterol of 70 mg/dL to 190 mg/dL. Outcome studies using this device suggested that those with a 10-year risk of ASCVD ≥ 7.5% should be considered for statin therapy.
However, application of this formula is limited by requiring the input of nine variables. For many outpatient primary care visits, not all these variables are available, thus limiting the application of the calculator.
This is a missed opportunity to discuss primary prevention interventions. CXRs are the most common imaging test done, and in the two hospitals, were available in the electronic medical record (EMR) in 20% of outpatients who fit the criteria for applying the ASCVD risk calculator. Studies have shown that in a CXR considered negative or normal, further information of value could be obtained by DL.
For example, DL analysis of CXRs has been shown to be able to detect early lung cancer and calculate coronary artery calcium content.
The Weiss et al study has shown that only 19% of patients who are candidates for estimating their risk of ASCVD had all the variables necessary to use the ASCVD calculator. Also, in those patients with all the variables to calculate their ASCVD risk, the CXR analysis had similar performance and added value to the ASCVD calculator.
Finally, in the 81% of patients in whom the data needed for the ASCVD calculator was not available, the risk of MACE was 1.5-fold higher for those deemed statin-eligible compared to those classified as ineligible by CXR.
However, the authors did not believe that CXR should be ordered for this purpose but rather that any existing CXR be used as an opportunistic analysis that can be conducted in the EMR.
There are limitations to the Weiss study. It is a retrospective, EMR-based study in two hospitals of mainly white, non-Hispanic individuals. Since this is hospital-based, the patients probably are at higher risk than a general population group would be. Also, these convolutional neural networks used to analyze the CXR basically are black boxes that are difficult to explain to patients, which may diminish enthusiasm for acting on their findings.
At this time, we need prospective trials to assess whether this DL CXR analysis can improve clinical decision-making and improve outcomes. However, it is clear that analyzing medical images likely will be one of the first forays of artificial intelligence into the practice of medicine.
A retrospective analysis of plain chest X-ray images in the medical record using deep learning in patients suitable for risk assessment for atherosclerotic cardiovascular disease (ASCVD) has shown similar results as the American College of Cardiology/American Heart Association ASCVD risk calculator for determining who is at sufficient risk to consider statin therapy.
Subscribe Now for Access
You have reached your article limit for the month. We hope you found our articles both enjoyable and insightful. For information on new subscriptions, product trials, alternative billing arrangements or group and site discounts please call 800-688-2421. We look forward to having you as a long-term member of the Relias Media community.