Artificial Intelligence Identification of Heart Failure with Preserved Ejection Fraction Substrate in Cardiac Surgery Patients: A Missed Opportunity

T. Liu¹, R. Nedadur¹, F. Ahmad², A. Baldrige³, J. D.. Thomas¹, A. Kline⁴, S. Shah¹, A. Narang¹, D. R. Johnston², P. M.. McCarthy^5
1Northwestern University Feinberg School of Medicine, Chicago, Illinois ²Northwestern Medicine, Chicago, Illinois ³Northwestern University, Chicago, Illinois ⁴Northwestern Medicine, Northwestern University, Chicago, Illinois ⁵Northwestern Medicine, Division of Cardiac Surgery, Chicago, Illinois

Purpose: Outcomes for cardiac surgery patients with heart failure (HF) with preserved ejection fraction (HFpEF) are not well understood. Adjudicating the presence of HFpEF is challenging in the presence of surgical pathology. We examined outcomes in patients with HFpEF echocardiographic findings identified by a commercially available, FDA-cleared artificial intelligence (AI) algorithm.

Methods: Between 2004-2022, adult cardiac surgery patients undergoing first-time, elective isolated aortic valve replacement (AVR) for aortic stenosis, mitral valve repair (MVr) for degenerative mitral regurgitation, or coronary artery bypass (CAB), with available preoperative transthoracic echocardiogram completed 90-days prior to operation were identified. Patients with reduced ejection fraction (LVEF < 50%) were classified as HF with reduced ejection fraction (HFrEF). Among patients with LVEF≥50%, the validated algorithm generates a probability score for HFpEF with higher probability associated with higher risk for adverse outcomes. Using the AI prediction score, we stratified patients into normal function (0-0.49), moderate probability HFpEF (0.5-0.74), and high probability HFpEF (≥0.75). Co-primary endpoints were all-cause operative and long-term mortality. Secondary endpoints were heart failure (HF) admission and atrial fibrillation (AF). We performed Kaplan-Meier analysis and examined risk for long-term outcomes with unadjusted and adjusted (age, sex, hypertension, renal failure, lung disease, stroke, and operation type) Cox proportional-hazard models.

Results: Among 1,882 patients (n=875 CAB, n=546 AVR, n=461 MVr), 86.6% (n=1629) had LVEF≥50%; of those 36.8% (n=599) were ≥0.75 probability HFpEF (Median ([IQR] LVEF: 60 [56,65]), 6.7% (n=109) were 0.5-0.74 probability HFpEF (LVEF: 61 [59,65]), and 56.5% (n=921) had normal function (LVEF: 63 [59,66]). The remaining 13.4% (n=253) patients were classified as HFrEF (LVEF: 40 [31,45]). Prevalence of high probability HFpEF varied by operation (CABG: 21.0%; AVR: 58.6%; MVr: 20.6%, p< 0.001). The STS predicted risk of mortality also varied by operation (Median ([IQR]; MVr: 0.4 [0.2,0.6]; AVR: 1.2 [0.7,2.3]; CABG: 0.7 [0.4,1.3], p< 0.001). High probability HFpEF was associated with higher operative mortality compared to normal function (1.3% vs. 0.3%, p=0.023) and 30-day readmission (12.8% vs. 6.7%, p< 0.001). Over median (range) of 5.8 (0-20.0) years of follow-up, high probability HFpEF had a greater incidence of mortality, HF admission, and AF than normal function (p < 0.01, Figure) with clinical outcomes in high probability HFpEF approaching the outcomes of patients with HFrEF. Risk-adjusted all-cause mortality was greater in high probability HFpEF compared to normal function (aHR: 1.84, CI: 1.35-2.52) (Table). High probability HFpEF had greater risk-adjusted HF admission and AF risk compared to normal.

Conclusion: Using AI, we can identify patients with high probability of HFpEF at high risk of early and late adverse events. Identifying this subgroup may enable surgical teams to improve short- and long-term outcomes through guideline-directed medical therapy and managed care pathways.

Identify the source of the funding for this research project: None