AI-based echo analyses best human analyses for COVID-19 mortality prediction
Click Here to Manage Email Alerts
In patients with COVID-19 who had an echocardiogram, analysis by artificial intelligence better predicted mortality than analysis by experts, according to the results of the WASE-COVID study.
For the study, presented at the American College of Cardiology Scientific Session, researchers enrolled 870 patients hospitalized with COVID-19 (mean age, 59 years; 44% women) who underwent transthoracic echocardiography.
In phase 1 of the study, to determine predictors of in-hospital mortality, the rate of in-hospital all-cause mortality was 21.6%, Federico M. Asch, MD, FASE, FACC, director of the Echocardiography Core Lab at MedStar Health Research Institute in Washington, D.C., and associate professor of medicine at Georgetown University, said during a presentation. Phase 1 was simultaneously published in the Journal of the American Society of Echocardiography.
“Myocardial injury has been linked with poor outcomes; therefore, an echocardiogram at admission may be a powerful tool to predict death,” Asch said.
In a multivariate analysis, left ventricular longitudinal strain was associated with in-hospital mortality (HR = 1.179; 95% CI, 1.045-1.358; P = .012), as were age, lactate dehydrogenase level, right ventricular free wall strain and previous lung disease, but LV ejection fraction was not, Asch said.
Phase 2 of the study, to compare AI software-based echocardiography analysis (Ultromics) with manual expert analysis, included echocardiograms from 476 patients for whom echocardiograms could be analyzed by both AI and manual expert methods. All-cause mortality through a mean follow-up of 230 days was 27.4%, Asch said.
“With reader-dependent technologies such as echo, fully automated, AI-based analysis should result in lower variability of results than those obtained from human reads,” Asch said during the presentation. “With increased interpretation consistency, it is foreseeable that the use of automated measurements could improve the capacity to predict outcomes.”
The experts and the software predicted mortality based on LV longitudinal strain and LVEF.
Variability was significantly larger for manual expert analysis than for AI analysis. Analysis by different operators was the main source of variance in the expert analyses (47.39% of variance for EF, 51.81% for longitudinal strain). AI software analysis had minimal variance, which was mostly due to selection of different video frames (6.3% variance for EF; 5.96% for longitudinal strain), Asch said.
The software outperformed the experts for prediction of in-hospital mortality (EF manual, OR = 0.985; 95% CI, 0.969-1.003; P = .083; EF software, OR = 0.97; 95% CI, 0.952-0.988; P = .001; longitudinal strain manual, OR = 1.035; 95% CI, 0.999-1.074; P = .058; longitudinal strain software, OR = 1.082; 95% CI, 1.035-1.132; P < .001) and follow-up mortality (EF manual, OR = 0.99; 95% CI, 0.975-1.005; P = .187; EF software, OR = 0.974; 95% CI, 0.956-0.991; P = .003; longitudinal strain manual, OR = 1.024; 95% CI, 0.991-1.059; P = .155; longitudinal strain software, OR = 1.06; 95% CI, 1.019-1.105; P = .004), according to the researchers.
“Automated quantification of left ventricular ejection fraction and left ventricular global longitudinal strain using artificial intelligence minimized variability,” Asch said. “AI-based LVEF and longitudinal strain analyses, but not manual, were significant predictors of in-hospital and follow-up mortality. AI analyses of echo could increase statistical power to predict outcomes, possibly requiring smaller sample sizes in clinical trials.”
Reference:
Perspective
Back to Top
Doreen DeFaria Yeh, MD
I found this study fascinating. Many of us in the echo community worry that AI may take our jobs one day.
We know LVEF is an important prognostic marker in many clinical situations. However, it is a crude marker of systolic function. In this study of echocardiographic predictors of COVID-19 outcomes, LV longitudinal strain was the parameter that was most predictive of in-hospital death, whereas LVEF was not significantly associated. Additionally, RV free wall strain was predictive of in-hospital death. That is not a surprise, given what we know about the development of acute pulmonary complications with COVID-19, which can cause acute RV strain. Abnormalities in the RV free wall may be expected in these patients as a marker of a more severe pulmonary condition. A significant portion of patients in this study (46%) were in the ICU with severe COVID-19 infection.
Looking at the manually obtained parameters vs. the AI-obtained parameters, AI was better at identifying patients who had subclinical abnormalities of ventricular function based on strain assessment. Importantly, AI had similar feasibility to manual contouring, with significantly less measurement variability. The AI-based analyses, but not the manual-based analyses, were significant predictors of in-hospital and follow-up mortality. This is a humbling study for echocardiographers because it tells us that a machine can make the echocardiographic measures and analyses that are most predictive of who might be at highest risk for death. Manual estimations of LVEF may vary significantly based on the quality of the 2D imaging, angulation, acoustic windows, etc, and this variability is certainly reduced with AI-based LV analyses.
This is an important paper as we use echocardiography every day in the assessment of our cardiac patients and critically ill patients. If some paraments may be more reliably and reproducibly assessed with AI, we should be move towards incorporating it into clinical practice. Establishing reproducibility of AI-derived LV analysis in other clinical contexts is critically important to evolving standard practice.
There will always be an important need for humans to oversee AI-interpreted data. However, it is important we do not close our eyes to the efficiencies and improved quality AI-derived analysis can bring in the appropriate context.
Collapse
Asch FM, et al. Featured Clinical Research II. Presented at: American College of Cardiology Scientific Session; May 15-17, 2021 (virtual meeting).
Read more about
Click Here to Manage Email Alerts