Comprehensive Summary
This study used a novel supervised machine-learning approach to classify aortic stenosis (AS) severity based on echocardiogram (ECHO) features in comparison to the traditional standard of care severity grading. An ECHO cohort (n = 1052) was used to train a machine-learning algorithm to label patients as having high or low severity. These classifications were then applied to the ECHO features of the original cohort, as well as to cohorts that underwent computed tomography (CT) (n=752) or cardiovascular magnetic resonance (CMR) (n=160). The algorithm-assigned severity label was then compared to the standard of care ECHO severity grading, the severity assessments based on the CT or CMR data, and the clinical outcomes of aortic valve replacement (AVR) and death. Almost all patients with severe graded AS and a majority of patients with mild/moderate or discordant AS were labeled by the classifier as high severity. In the CT cohort, a majority of the patients with severe calcium scores were labeled as high severity. In the CMR cohort, myocardial volume, fibrosis volume, other predictors of severe AS, and cardiac biomarkers of heart failure were all significantly higher in patients labeled as having high severity. In the ECHO and CMR cohorts, patients labeled as having high severity needed AVR faster compared to those labeled low severity, even if they received a traditional label of mild/moderate or discordant. In the ECHO cohort, patients labeled as having high severity progressed to death faster, regardless of AVR treatment. These results showed that the machine-learning model could consistently classify severe AS ECHO features into high severity and reclassify mild/moderate and discordant features to make predictions about future clinical events.
Outcomes and Implications
The standard of care severity grading of AS based on ECHO features has high levels of diagnostic uncertainty, especially in discordant results. The presentation of AS in the valve is very diverse, and patients with discordant ECHO results often need more testing and experience a delayed clinical response. Tests that focus on the left ventricular (LV) myocardial response to pressure overload, such as CT or CMR, are more accurate but also more expensive, not widely accessible, and expose patients to harmful radiation or intravascular contrast agents. This research provides an alternative method of a machine-learning classifier that is accurate, simpler, and has been made publicly available by the authors. It can categorize patients based on ECHO features alone, but can optimize the timing of AVR and make predictions about future clinical events. As this study is only observational, more clinical research is needed before implementation, but by expanding the use of machine learning in AS severity grading, clinical decision-making and patient outcomes may improve.