Aortic stenosis (AS) is a common valvular heart disease which may be fatal without treatment [1]. A significant challenge lies in risk stratifying individuals with severe low-gradient (LG) AS, a subset characterized by reduced transvalvular pressure gradient. These patients are at a higher risk vs. high-gradient AS, making traditional risk stratification models less effective [2]. Patients with severe LG AS have varied prognoses depending on factors such as subtype and underlying etiology. Although aortic valve intervention (AVR) improves survival, its benefit is less pronounced in this group, emphasizing the need for personalized risk assessment tools [1]. Machine learning (ML) offers promise in improving risk stratification for AS [3]. ML models can potentially outperform traditional risk models by utilizing nonlinear signals to make accurate predictions [4]. However, the application of ML models in severe LG AS has been incompletely explored [5,6]. This study aimed to evaluate the utility of a validated ML model, originally designed for general AS risk prediction, to identify higher-risk individuals with severe LG AS. We hypothesized that the same features used in general AS risk stratification could effectively sub-phenotype patients with severe LG AS, offering new insights into this complex disease.
The study utilized TTE report data at Beth Israel Deaconess Medical Center linked to Medicare claims, 2003–2020, including comprehensive death information. The study was approved by the Institutional Review Board, with patient consent waived. Only patients ≥ 65 years with native severe LG AS (defined as having an aortic valve area < 1.0 cm2 and either a peak aortic velocity < 4 m/s or an aortic mean gradient < 40 mmHg) for whom stress echocardiography concluded truly severe aortic AS were included. A previously published ML model designed for general AS risk stratification was used [6]. The model has been made publicly available (https://as-gps.herokuapp.com/) and utilizes five variables from TTE reports: aortic valve peak velocity, mean gradient, stroke volume index, left ventricular ejection fraction (LVEF), and indexed aortic valve area. The model predicts risk groups and probabilities using a cluster-then-predict approach. The primary endpoint was all-cause mortality within five years of the index TTE, with secondary outcomes including receipt of AVR. Demographic, anthropometric, and echocardiographic parameters were recorded, along with comorbidities determined using Medicare Chronic Conditions Warehouse data. Kaplan-Meier curves and Cox proportional hazards models were used to assess mortality risk, both unadjusted and adjusted for age, gender, LVEF, coronary artery disease, heart failure, and mitral regurgitation severity, censoring at last follow-up date.
Of 3,604 individuals with native AS, 521 with severe LG AS were included in the study. The ML model identified 60.1% as high-risk and 39.9% as low-risk. High-risk frequently male, had coronary artery disease, heart failure, ≥ moderate mitral regurgitation, and had lower ejection fraction (Table; LVEF; all p < 0.05). At 5-years, 387 (74.3%) participants died, 241 (77.0%) in the higher-risk group and 146 (70.2%) in lower-risk (Figure; HR 1.32, 95% CI 1.08–1.62, p = 0.008). Overall, 100 (19.2%) patients underwent AVR, including 71 (22.7%) higher-risk and 29 (13.9%) lower-risk (HR 1.21, 95% CI 0.86–1.70, p = 0.28). The ML model correctly stratified unadjusted mortality risk among those receiving (HR 2.66, 95% CI 1.11–6.41, p = 0.03) and not receiving AVR (higher vs. lower-risk, HR 1.56, 95% CI 1.25–1.94, p < 0.001). After adjustment for age, gender, LVEF, coronary artery disease, heart failure, and mitral regurgitation severity, risk-category was significantly associated with AVR receipt (adjusted HR 1.62, 95% CI 1.03–2.54, p = 0.04) but not mortality (adjusted HR 0.89, 0.68–1.18, p = 0.43). Risk prediction was consistent across subgroups (Supplemental Figure) and incrementally improved prediction of death or AVR compared to cardiac damage staging (Supplemental Table) [7]
Table:
Baseline characteristics of individuals with model output and low-gradient severe AS
| Overall (N = 521) |
Higher-Risk AS (N = 313) |
Lower-Risk AS (N = 208) |
p-value | |
|---|---|---|---|---|
| Demographic Profile | ||||
| Age, years | 83.7 ± 7.6 | 83.2 ± 7.7 | 84.5 ± 7.4 | 0.047 |
| Women (%) | 270 (51.8%) | 123 (39.3%) | 147 (70.7%) | < 0.001 |
| White, % | 494 (94.8%) | 301 (96.2%) | 193 (92.8%) | 0.24 |
| Black, % | 14 (2.7%) | < 11 | < 11 | 0.24 |
| Other, % | 13 (2.5%) | < 11 | < 11 | 0.24 |
| Clinical Profile | ||||
| Inpatient, % | 392 (75.2%) | 247 (78.9%) | 145 (69.7%) | 0.02 |
| Body mass index, kg/m2 | 26.2 ± 5.3 | 26.7 ± 5.4 | 25.5 ± 5.0 | 0.01 |
| Systolic / Diastolic blood pressure, mmHg | 128.2 ± 23.3 / 65.6 ± 14.2 | 123.5 ± 22.3 / 65.6 ± 13.6 | 135.3 ± 23.0 / 65.7 ± 15.1 | < 0.001 / 0.95 |
| Heart rate, bpm | 75.0 ± 15.2 | 75.9 ± 16.2 | 73.6 ± 13.6 | 0.08 |
| eGFR, L/min/1.73m - median (IQR) | 76.8 (0–203.0) | 0 (0–173.6) | 141.1 (0–220.6) | < 0.001 |
| NT-proBNP, pg/ml – median (IQR) | 6983 (1893–14048) | 2282 (837–7195) | 8293 (3801–16947) | < 0.001 |
| Past Medical History | ||||
| Type II Diabetes, % | 171 (32.8%) | 109 (34.8%) | 62 (29.8%) | 0.25 |
| Hypertension, % | 397 (76.2%) | 239 (76.4%) | 158 (76.0%) | 0.92 |
| Coronary artery disease, % | 383 (73.5%) | 245 (78.3%) | 138 (66.4%) | 0.003 |
| Heart failure, % | 351 (67.4%) | 233 (74.4%) | 118 (56.7%) | < 0.001 |
| Aortic Valve Profile | ||||
| Peak velocity, m/s | 3.3 ± 0.5 | 3.4 ± 0.5 | 3.1 ± 0.4 | < 0.001 |
| Mean gradient, mmHg | 25.2 ± 7.5 | 27.2 ± 7.4 | 22.1 ± 6.4 | < 0.001 |
| Aortic valve area, cm2 | 0.86 ± 0.14 | 0.82 ± 0.15 | 0.93 ± 0.09 | < 0.001 |
| Right Ventricular Function & Dimensions | ||||
| Peak TR velocity, m/s | 3.0 ± 0.5 | 3.0 ± 0.5 | 3.0 ± 0.5 | 0.16 |
| ≥ Moderate TR | 132 (25.3%) | 86 (27.5%) | 46 (22.1%) | 0.18 |
| Left Atrial and Ventricular Function & Dimensions | ||||
| Left atrial volume index, mL/m2 | 36.5 ± 11.5 | 37.9 ± 11.9 | 35.0 ± 10.9 | 0.27 |
| LV end diastolic dimension, cm | 4.5 ± 0.8 | 4.8 ± 0.8 | 4.1 ± 0.6 | < 0.001 |
| LV end systolic dimension, cm | 3.0 ± 1.0 | 3.4 ± 1.0 | 2.5 ± 0.5 | < 0.001 |
| LV ejection fraction, % | 55.1 ± 19.5 | 45.3 ± 17.3 | 69.9 ± 11.7 | < 0.001 |
| Transmitral E/e’ ratio | 15.8 ± 5.9 | 16.3 ± 6.1 | 14.7 ± 5.5 | 0.28 |
| Transmitral E/A ratio | 1.4 ± 0.9 | 1.4 ± 1.0 | 1.2 ± 0.7 | 0.18 |
| Stroke volume index, mL/m2 | 34.2 ± 9.8 | 31.8 ± 9.6 | 37.7 ± 9.1 | < 0.001 |
| ≥ Moderate Mitral Regurgitation, % | 133 (25.6%) | 97 (31.0%) | 36 (17.3%) | < 0.001 |
| Pharmacotherapy | ||||
| Anticoagulant, % | 67 (12.9%) | 45 (14.4%) | 22 (10.6%) | 0.23 |
| Diuretic % | 137 (26.3%) | 87 (27.8%) | 50 (24.0%) | 0.36 |
| Neurohormonal antagonists, % | 131 (25.1%) | 74 (23.6%) | 575 (27.4%) | 0.35 |
| Anti-platelet, % | 143 (27.4%) | 84 (26.8%) | 59 (28.4%) | 0.76 |
| Anti-arrhythmic, % | 168 (32.2%) | 103 (32.9%) | 65 (31.3%) | 0.70 |
| Beta-blocker, % | 163 (31.3%) | 98 (31.3%) | 65 (31.3%) | > 0.99 |
| History of Coronary Intervention | ||||
| PCI, % | 27 (5.2%) | 18 (5.8%) | < 11 | 0.55 |
| CABG, % | 23 (4.4%) | 19 (6.1%) | < 11 | 0.03 |
This table outlines the demographic, clinical, echocardiographic, and treatment characteristics for each category of AS, classified by artificial intelligence-decision support algorithm. Data are presented as means ± standard deviations or counts with percentages. The number of individuals with complete data for each variable is provided. CABG = coronary artery bypass grafting, eGFR = estimated glomerular filtration rate, LV = left ventricular, N = number of patients, NT-proBNP = N-terminal probrain natriuretic peptide, PCI = percutaneous coronary intervention, TR = tricuspid regurgitation. Cell values < 11 are suppressed from publication as per Medicare data use agreements.
Figure: Kaplan-Meier Curve Evaluating Time to Death by AI-Risk Category and Receipt of Aortic Valve Replacement Among Individuals with Severe Low-gradient Aortic Stenosis.
Displayed is a Kaplan Meier curve demonstrating the proportion of individuals without an event (death; y-axis) according to time in years (x-axis) from the index echocardiogram by machine-learning (ML) designated risk group and receipt of aortic valve replacement (AVR). The red curve indicates individuals determined to be higher-risk by the ML algorithm who received an AVR in follow-up, the green curve indicates individuals determined to be lower-risk who received an AVR, the blue curve indicates individuals at higher-risk who did not receive an AVR, and the orange curve indicates individuals at lower-risk who did not receive an AVR. Numbers in the risk set are provided below at 1-year intervals. The log-rank p-value for comparison across groups was <0.001.
Among those meeting inclusion, this study demonstrates that a ML model designed for general AS can effectively stratify risk in patients with severe LG AS. Despite being agnostic to clinical conditions or outcomes, the model identified individuals at higher risk of death on an unadjusted basis, regardless of AVR status, perhaps partially due to inclusion of LG AS patients among those in the derivation cohort. While the magnitude of risk stratification was greater among those who underwent AVR, these findings support the potential of ML models to enhance prognostication and guide clinical decisions in this complex subgroup of AS patients. In addition to identifying if this ML model could stratify benefit of AVR in randomized trials, prospective testing is necessary to determine its utility in real-world clinical settings. Moreover, further research is needed to identify the optimal probability cutoff for risk in this population.
Among those meeting inclusion, a ML model effectively identified high-risk patients with severe LG AS, suggesting its potential use in clinical decision support. However, further prospective studies are required to determine its utility in real-world clinical settings.
Supplementary Material
Funding:
Dr. Strom reports research grants from the National Institutes of Health (1R01HL169517, 1R01AG063937).
Disclosures:
Dr. Strom reports research grants from the National Heart, Lung, and Blood Institute (1R01HL169517) and National Institute of Aging (1R01AG063937), Anumana, , Philips Healthcare, and Bracco Diagnostics; consulting for Bracco Diagnostics, Edwards Lifesciences, Philips Healthcare, General Electric Healthcare, and EVERSANA Lifesciences, and is a member of the scientific advisory boards for Ultromics, HeartSciences, EchoIQ, Bristol Myers Squibb, and the data safety monitoring board for Pfizer. Dr. Sengupta has received research grants from National Science Foundation (Grant #2125872), HeartSciences, RCE Technologies, MindMics Inc., Butterfly Inc., Us2ai (research software support) and serves as a consultant to HeartSciences and RCE Technologies. Dr. Yanamala is a member of the advisory board for TurnKey TechStart and Research Spark Hub.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
References
- 1.Strange G, Scalia GM, Playford D, et al. Uncovering the treatable burden of severe aortic stenosis in Australia: current and future projections within an ageing population. BMC Health Serv Res. 2021;21:790. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Vogelgesang A, Hasenfuss G, Jacobshagen C. Low-flow/low-gradient aortic stenosis-Still a diagnostic and therapeutic challenge. Clin Cardiol. 2017;40:654–659. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ito S, Oh JK. Aortic Stenosis: New Insights in Diagnosis, Treatment, and Prevention. Korean Circ J. 2022;52:721–736. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Sengupta PP, Chandrashekhar Y. From Conventional Deep Learning to GPT: AI’s Emergent Power for Cardiac Imaging. JACC Cardiovasc Imaging. 2023;16:1129–1131. [DOI] [PubMed] [Google Scholar]
- 5.Chokesuwattanaskul R, Petchlorlian A, Lertsanguansinchai P, et al. Machine Learning-Based Predictive Model of Aortic Valve Replacement Modality Selection in Severe Aortic Stenosis Patients. Med Sci. 2024;12:3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Sengupta PP, Shrestha S, Kagiyama N, et al. A Machine-Learning Framework to Identify Distinct Phenotypes of Aortic Stenosis Severity. JACC Cardiovasc Imaging. 2021;14:1707–1720. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Généreux P, Pibarot P, Redfors B, et al. Staging classification of aortic stenosis based on the extent of cardiac damage. Eur Heart J. 2017;38:3351–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.