Conference symposium

Background

• Detection of HFpEF involves integration of multiple imaging and clinical features which are often discordant or indeterminate
• Hypothesis: Applying artificial intelligence to detect HFpEF from a single apical four-chamber transthoracic echo video clip; accuracy will be better then clinical scores

Methods

• 3D convolution neural network was developed and rained on A4C video clips
  • HFpEF: diagnosis of HF, EF ≥50%, and echocardiographic evidence of increased filling pressure , n=3785
  • Without HFpEF: No diagnosis of HF, EF ≥50%, normal filling pressure, n=2971
  • Validation holdout 6.8%
• Softmax activation function to calculate a value 0-1, mapped to HFpEF or no HFpEF; non-diagnostic output generated based on model uncertainty
• Compared observed sensitivity and specificity in an independent testing data set to average reported in the literature (74% sensitivity, 65% specificity)
• To demonstrate a 5% increase from the benchmark, and allowing for ~20% nonmagnetic outcomes, 1048 patients were required in an independent testing data set.
• Multicentre independent retrospective data from Mayo Clinic Health system used for testing.
  • Cases and controls matched for sex and year of echocardiogram and attempts were made to match for age (72.4 yrs vas 4.6 yrs)
  • Up Sampling of non-white and Hispanic populations OT ensure the model would work in a more diverse population.

Results

Testing in 1284 patients

• Non-diagnostic in 94 (7.3%)  
• Sensitivity 87.8%
• Specificity 81.9%
• During median follow up of 2.3 [0.5-5.6 year] 444(34.6%) patients died
• Age-adjusted mortality was higher in pts classified as HFpEF by AI

Conclusion

A novel AI HFpEF model based on a single, routinely acquired echocardiographic video demonstrated excellent discrimination of patients with versus without HFpEF, more often than the HFA-PEFF or H2FPEF score and identified patients with higher mortality.