A team of researchers at UCSF developed a technology called CathEF that uses artificial intelligence (AI) to assess how well the heart pumps blood during a common heart test.
Usually, when doctors suspect someone might be at risk of a heart attack, they carry out a test known as a coronary angiography. This procedure checks if the heart's blood vessels are blocked or narrowed, which could indicate a risk of heart attack.
To get a better understanding of how well the heart is pumping blood, doctors can also do another procedure called a left ventriculography. This involves inserting a tube into the left side of the heart and injecting a special dye. However, this extra procedure can pose additional risks to the patient.
The team's invention, CathEF, can estimate how well the heart is pumping by analyzing videos taken during a standard coronary angiography. This means doctors could get this information without needing the extra procedure, which could make the angiography safer.
Our work demonstrates that video-based AI can achieve fully-automated and accurate estimation of LVEF from standard routinely-obtained coronary angiograms of the left coronary artery. This provides an opportunity to estimate LVEF during nearly every angiogram in a noninvasive, risk-free manner.
To develop CathEF, we first analyzed real-world coronary angiograms using CathAI—a pipeline of multiple deep neural network algorithms—to automatically identify angiogram videos containing the left coronary artery (LCA) as their primary anatomic structure.
.png)
Angiograms were paired with a corresponding transthoracic echocardiogram (TTE) assessment of LVEF performed either 3 months before or 1 month after the angiogram.
Our final UCSF dataset consisted of 26,087 LCA videos derived from 3,960 coronary angiograms and 3,404 distinct patients. This was then split into training, development and testing datasets to develop and internally validate CathEF.
CathEF demonstrated strong performance to estimate low LVEF ≤40% in the UCSF test dataset. We then additionally validated it in an external validation dataset in real-world angiograms from the University of Ottawa Heart Institute, showing similarly strong performance.
In the UCSF test dataset (n=813), CathEF showed strong discrimination for LVEF ≤40% with an area under the receiver operating characteristic curve (AUC) of 0.911 (95% CI 0.887-0.934). We also tested discrimination for the cutoff of LVEF ≤50% and AUC=0.879 (95% CI 0.852-0.907). There were 22.7 greater odds of reduced LVEF in those that CathEF predicted LVEF ≤40%. Sensivitiy and specificity were 83.9% and 81.3%, respectively.
CathEF also estimated continuous LVEF percentages, comparing well against TTE LVEF. The median absolute error of CathEF versus TTE-LVEF was 8.5% (95% CI 8.1%-9.0%), and it was 3.7% (95% CI 2.7%-9.3%) when averaging CathEF's predictions across a patient's angiographic study. Intraclass correlation coefficient was 0.77.
CathEF remained consistent across sex, BMI, presence of obstructive coronary artery disease, left ventricular hypertrophy and reduced kidney function. This includes many patients for whom the additional contrast from left ventriculography can be harmful. CathEF performed similarly for patients presenting urgently for ACS as for non-ACS.
To examine external generalizability, we validated CathAI in real-world angiograms from the University of Ottawa Heart Institute, a separate healthcare institution in a different country.
In 4,471 LCA videos derived from 776 angiograms from 744 patients paired with TTE-LVEF, CathEF achieved an AUC of 0.906 (95% CI: 0.881-0.931) to identify LVEF ≤40%. There were 27.3 greater odds of reduced LVEF in those that CathEF predicted LVEF ≤40% in the UOHI dataset. Sensitivity and specificity for LVEF ≤40% were 77.9% and 88.6%, respectively.
When predicting continuous LVEF percentages, median absolute error was 7.0% compared to TTE-LVEF.
We used the guided GradCAM algorithm explainability techqniues to better understand how CathEF predicts LVEF from angiograms.
GradCAM consistently highlighted epicardial coronary arteries during cardiac systole, and largely not during diastole. This suggests that CathEF likely identifies patterns in coronary artery blood flow or movement during systole to predict LVEF.
Our work demonstrates that CathEF, a video-based neural network, can accurately estimate LVEF, a key measurement of cardiac pumping function, providing valuable information from standard coronary angiograms that is usually not accessible to clinicians, with no additional procedures or risk.
CathEF offers the opportunity to obtain real-time LVEF estimates from routine angiograms with no added risk or procedures.
CathEF could potentially change the way heart tests are done by providing real-time information about the heart's pumping ability during a standard angiography. This could be particularly helpful for patients who need urgent care, such as those suspected of having a heart attack. It might also benefit patients who are at risk of kidney disease, as the contrast dye used in the extra procedure can be harmful to them.
