AI uses imaging results to ID high-risk TAVR patients with speed, accuracy

Researchers have developed a new way to identify patients who may face an especially high risk of death after transcatheter aortic valve replacement (TAVR) based on unprocessed CT images. The group shared its findings in Scientific Reports, noting that artificial intelligence (AI) helped them do in seconds what typically takes trained specialists 10 or 15 minutes.[1]

“Before an intervention, patients undergo a comprehensive assessment based on various factors, including their health status and medical history,” wrote first author David Brüggemann, PhD, a specialist with the Computer Vision Laboratory in Switzerland, and colleagues. “An important component of this assessment is a CT scan of the chest and abdomen … The CT image is analyzed qualitatively or quantitatively, with manually extracted measurements that are believed to be linked to the procedural outcome. However, the need for a radiologist to extract these measurements makes the preprocedural assessment more time-consuming and expensive.”

Brüggemann et al. designed advanced AI algorithms capable of using unprocessed, preprocedural CT images and baseline patient characteristics to determine a patient’s risk profile prior to TAVR. The AI can even adapt in cases when certain data are missing, reviewing the available information to provide a helpful estimate of the patient’s post-TAVR outcomes.

The group used data from nearly 1,500 TAVR patients treated at a single center to build its algorithms. The AI was able to localize five landmarks on the centerline of the aorta in unprocessed CT images thanks to the combination of a “reinforcement learning-based approach” and a “regression-based approach.” The final product was a model capable of extracting key measurements from CT images with significant accuracy and then using those measurements to make an effective prediction of the patient’s mortality risk.

“The results … confirm that our automatic feature extraction can replace manually extracted image measurements without forfeiting prediction accuracy, which is an important and surprising point,” the authors wrote. “The manually extracted features are complex measurements that integrate clinical experience and medical knowledge. However, extracting such features for every patient can create a bottleneck during preprocedural patient assessment: Extracting the measurements manually from a single image takes an expert radiologist between 10 and 15 minutes depending on calcification severity. Our model delivers a prediction within 5 to 20 seconds on a consumer computer—depending on the number of missing variables—allowing patient assessment with minimal manual labor.”

The new AI model’s final area under the ROC curve (AUROC) was 0.725 for predicting all-cause mortality during post-TAVR follow-up. This is similar to the work of expert radiologists, the authors noted, and the comparable performances suggest the proposed model could go a long way toward helping cardiologists identify high-risk TAVR patients.

Brüggemann and colleagues did emphasize that there are certain limitations to their work. For example, they primarily focused on post-TAVR mortality over the course of their analysis, but there are other potential outcomes that could have been explored.

“A more comprehensive measure of TAVR success should incorporate not only mortality, but also the patient’s quality of life and functional level after the procedure,” they wrote. “This could provide a clinically more useful estimate of TAVR risk. Additionally, there may be other important predictors of TAVR risk that were not collected in the examined TAVR registry, e.g., atrial fibrillation. However, our model does not necessitate structural modifications to accommodate additional variables or predict different outcomes. If a TAVR registry with this information becomes available, our model could be fit analogously.”

Read the full study in Scientific Reports here.

Michael Walter
Michael Walter, Managing Editor

Michael has more than 16 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.

Around the web

Ron Blankstein, MD, professor of radiology, Harvard Medical School, explains the use of artificial intelligence to detect heart disease in non-cardiac CT exams.

Eleven medical societies have signed on to a consensus statement aimed at standardizing imaging for suspected cardiovascular infections.

Kate Hanneman, MD, explains why many vendors and hospitals want to lower radiology's impact on the environment. "Taking steps to reduce the carbon footprint in healthcare isn’t just an opportunity," she said. "It’s also a responsibility."

Trimed Popup
Trimed Popup