Skip to content

Deep Learning for Detection of Preclinical Heart Failure in the Multi-Ethnic Study of Atherosclerosis

  • | By Ultromics

Amogh Karnik1, Lauren Nelson1, Ben Freed1, Ross Upton2, Ashley Akerman2, Ambarish Pandey3, Alain Bertoni4, Sanjiv Shah1, Faraz Ahmad1

1Northwestern University Feinberg School of Medicine 2Ultromics Ltd 3University of Texas Southwestern School of Medicine 4Wake Forest University School of Medicine

 

ACC_2025_Deep Learning for Detection of Preclinical Heart Failure in the Multi-Ethnic Study of Atherosclerosis

See full-poster

Background

  • MESA is a highly diverse, community based study of adults (free of cardiovascular disease) in the United States followed since enrollment in 2000-2002.
  • Exam 6 took place between 2016-2018, and involved diagnostic testing including comprehensive echocardiography, six minute walk testing, and the Kansas City Cardiomyopathy Questionnaire
  • In this analysis - we sought to evaluate the performance of the EchoGo HF classifier within this study population for detection of heart failure

Methods

  • N = 3021 participants MESA Exam 6; 57 were excluded due to prior HF
  • N = 2432 participants completed all KCCQ, 6MWT, and comprehensive echocardiography
  • 37,830 individual clips analyzed for predictions
  • For each participant, the HF prediction was based on the majority vote of predictions for all obtained clips
  • Each prediction was assigned the median probability of the clips that corresponded with the majority
  • Participants classified as having HF Detected were divided into two subgroups: probability ranging from 0.75 - 1, and 0.50 - 0.74

Results

  • Mean age was higher in the HF Detected groups
  • BMI was higher in the HF Detected groups
  • Distribution between ethnicities was similar in all groups
  • Hypertension, diabetes, and current/former smoking status were all more common in those with higher HF probabilities
  • NT-proBNP was higher in those with higher HF probabilities
  • Total 6MWT distance was lower in those with higher HF probabilities
  • KCCQ clinical summary scores were lower in those with higher HF probabilities
  • Multivariate linear regression modeling revealed a statistically significant relationship between HF probability score, 6MWT distance, and KCCQ clinical summary score

clinical para

 

table 2

 

regression analysis


Discussion

After adjusting for key clinical and echocardiographic parameters, the AI predicted HF probability was significantly associated with lower 6MWT distance and lower clinical summary summary scores.

In patients without known clinical heart failure, higher AI predicted HF probability was associated with clinically evident changes both in functional capacity and assessment of symptoms.