Artificial Intelligence Versus Sonographer Echocardiogram Analysis and Reporting in Patients With Heart Failure

Heart Failure

Echocardiogram, Artificial Intelligence, Alogrithm, Echocardiological image analysis

This is a non-inferiority, three-year, multicenter, double-blinded randomized controlled study of an AI versus experienced sonographer echocardiogram analysis in HF patients. Consecutive patients presented for echocardiogram examination with new or worsening HF symptom and positive HF blood markers will be recruited. A target of 514 patients will be randomized 1:1 to receive either AI or sonographer echocardiogram analysis. The primary endpoint of diagnostic accuracy is the complete agreement of disease grading with an experienced cardiologist (American Society of Echocardiography level III) using a standardized grading chart. Important secondary endpoints include the time used for echocardiogram report drafting and report endorsement, 6-month heart failure symptom and hospitalization, and the cost-effectiveness of AI to increase echocardiogram service. Clinical, biochemical and echocardiographic predictors of worsening of heart failure and hospitalization will be identified.

Background Unmet Need for Streamlined Echocardiogram Algorithm Heart failure (HF) is a global pandemic affecting more than 64 million people in the world. In Hong Kong, the prevalence of HF is estimated to be 2-3% with a steep rise of new onset HF hospitalization in the older age group. The estimated annual worldwide economic burden of HF was 108 billion United States dollars, with direct costs to healthcare systems accounted for 60% and indirect costs to society driven by premature mortality, morbidity and lost productivity accounted for the remaining 40%. Timely diagnosis of HF etiology with early appropriate treatment are critical to reduce HF hospitalization and mortality. While HF with reduced ejection fraction (HFrEF) and preserved ejection fraction (HFpEF) requires different guideline directed medical therapy (GDMT), HF patients with severe valvular heart disease requires interventional treatment. Echocardiogram (cardiac ultrasound) is the key diagnosticmodality to phenotype HF and to guide subsequent appropriate treatment. Access to echocardiogram in Asia Pacific is severely limited (e.g. average waiting time in Hong Kong for routine echocardiogram is 12-18 months), which results in delay in appropriate treatment and hence poor outcomes. While image acquisition is easier to teach, analysis in echocardiogram is time consuming and requires years of training to become proficient, and yet has significant inter-observer variability. Therefore, there is a shortage of fully trained sonographers globally. A streamlined echocardiogram analysis pathway that can enhance the efficiency while improving the diagnostic accuracy of HF etiology is appealing. Emerging role of Artificial Intelligence in Echocardiogram Artificial intelligence (AI) has emerged as a useful tool with the potential to enhance cardiovascular care including in disease diagnosis, treatment guidance and outcome prediction. Collaborator of this study, David Ouyang et al., has developed machine learning algorithm for fully automated assessment of left ventricular ejection function (LVEF), aortic valve stenosis (AS) and mitral valve regurgitation (MR), with similar accuracy compared to manual analysis by experienced sonographers with reference to cardiologists ("gold standard"). Similar works has also been done by other teams. However, most of these validation studies are conducted based on retrospective echocardiogram cohort. Besides, there can be bias when a different sonographer than the scanning sonographer interprets the images, and that potentially compromised the real-life diagnostic accuracy of sonographers. Local Heart Failure Data and Application Artificial Intelligence in Echocardiogram Studies from our team has demonstrated that early diagnosis and intensified HF GDMT can reduce HF hospitalization from 13.1% to 8.6% (Hazard ratio = 0.65, p\<0.01). Besides, a strong association of 30-day unplanned HF hospitalization with severe valvular heart disease, mostly AS or MR, was found (Odd ratio =72.04, p=0.03). This implies that early phenotyping the mechanism of HF is important. From our unpublished pilot data of patients presented with HF symptom, echocardiogram image acquisition took only 54.2% of the total echocardiogram process time while the remaining were used for analysis by sonographer. When compared, AI used a significantly shorter time for echocardiogram analysis (324 seconds vs 1057 seconds, p\<0.01), with a 91.6% agreement rate on LVEF grading and severity of AS and MR. However, this pilot data was collected retrospectively, and the sample size was small. Therefore, it remains unclear whether AI is as accurate and more efficient than experienced sonographers in analyzing multiple possible echocardiogram abnormalities that can interact with each other for HF patients. Moreover, whether the addition of AI analysis will affect the final grading by cardiologists has not been studied. In this research project proposal, Investigator aim to assess whether a tailored AI echocardiogram analysis and reporting system is as accurate as an experienced sonographer in HF patients by conducting a multicenter double-blinded randomized controlled study.

No related papers found yet.

Hong Kong