Key Points
- Aortic stenosis, despite being life-threatening, is often underdiagnosed. It is understood that when the disease is diagnosed, it does remain undertreated as well.
- An artificial-intelligence algorithm was developed with the intent of being a decision support algorithm, both to identify patients with aortic stenosis, and stratify those patients by risk.
- Using the National Echo Database of Australia, containing 1 million echocardiograms, the algorithm was able to identify high-risk phenotypes, correctly identifying all of the patients diagnosed with severe aortic stenosis, and flagging others who were likely to have the disease but failed to meet criteria.
Severe, symptomatic aortic stenosis (AS) is a debilitating, life-
threatening condition. By the time symptoms develop, patients have up to a 50% mortality rate within two years if left untreated. Due to the increasing age of our patient populations, the prevalence of severe aortic stenosis is also increasing. Due to such high-stakes, it is imperative that patients with severe aortic stenosis are diagnosed earlier in the disease process, so that therapies may be initiated when indicated. Despite this, 30-54% of patients with severe AS remain untreated, even after diagnosis by echocardiography. The investigators of AI-ENHANCED-AS sought to improve the sensitivity of standard echocardiography by using an AI algorithm to increase the detection of severe aortic stenosis. In a Hot Line Session at the 2022 European Society of Cardiology Congress in Barcelona, Dr. Geoffrey Strange (Notre Dame University, Australia) presented the findings of their study.
The study investigators used a proprietary AI-Decision Support Algorithm that was trained using the National Echo Database of Australia, with over 1,000,000 echocardiograms linked to mortality data. 400 input nodes, and 4 hidden layers of over 2000 nodes were used in the neural network, with the “golden rule” for severe aortic stenosis being an aortic valve area (AVA) of ≤ 1.0 cm2. The algorithm was instructed on the proper definitions of aortic stenosis severity based on the current guidelines to ensure aortic stenosis was detected. After learning from 70% of the database, the algorithm was put to the test using the remaining 30%, comprising echocardiograms from 184,301 patients.
The AI Decision Support Algorithm (AI-DSA) identified 2,606 (1.4%) patients with a moderate-to-severe risk phenotype and 4,622 (2.5%) with a severe phenotype. 77.2% of patients in the severe group met the current guideline criteria for severe aortic stenosis. The 5-year mortality of patients in these groups was then assessed: moderate to severe phenotype was associated with a 56.2% mortality rate and severe was associated with a 67.9% mortality rate. The reference group, those without either phenotype, also had a higher-than-expected mortality rate of 22.9%. Despite this, the age and sex-adjusted odds ratio for all cause mortality was high for either group: OR 1.82 for the moderate-to-severe phenotype (95% confidence interval [CI] 1.63–2.02) and 2.80 for the severe phenotype (95% CI 2.57–3.06). Within the severe phenotype are those patients who met current guidelines for severe aortic stenosis, and those who did not. Of the patients who met guideline criteria, the five-year mortality rate was higher, at 69.1%. Characteristics that were observed in patients who did not meet criteria but still had a high mortality risk were large left atria, thickened left ventricles, and the presence of diastolic dysfunction.
Dr. Catherine Otto provided a summary of the findings, and stated that the more striking result was he amount of echocardiograms that did not report the fundamental elements required to diagnose the severity of aortic stenosis. She additionally stressed the importance of individualizing echocardiographic data to the patients at hand, stating “guidelines are not meant to provide strict cutoffs, but rather to steer one in the correct direction with regard to patient care. From a quality standpoint, all echo reports should include aortic valve data in the summary, with a verbal interpretation of how severe the reader feels the valve to be”. Dr. Strange agreed, and added that if there are discrepancies in the data, other echo parameters should be used to enhance our diagnostic acumen.
