Log in to MyACC
Machine Learning to Optimize Echocardiographic Follow-Up of Aortic Stenosis
Quick Takes
- Using echocardiograms from patients with mild or moderate aortic stenosis (AS) defined using transaortic Vmax, a machine learning model using age and nine echocardiographic variables was used to predict the relative rate of progression to severe AS.
- Predicting slower rates of progression predicted by the machine learning model could have resulted in performance of fewer follow-up echocardiograms with resulting cost savings compared to current European and American guideline recommendations.
- The study is limited by its characterization of AS severity, the use of echocardiograms rather than patients as an endpoint, and a surprisingly high rate of progression to severe AS during the relatively short follow-up interval.
Study Questions:
Can machine learning help optimize the frequency of echocardiographic surveillance of patients with aortic stenosis (AS)?
Methods:
Demographic and echocardiographic data from 4,633 echocardiograms in 1,638 patients with mild (transaortic peak velocity [Vmax] 2.0-2.9 m/s) or moderate AS (Vmax 3.0-3.9 m/s) and ≥1 follow-up echocardiogram at a single tertiary hospital in Spain were used to create a machine learning model to try to predict whether severe AS (Vmax ≥4.0 m/s) would develop in 1, 2, or 3 years. An external cohort of 4,531 echocardiograms in 1,533 patients with mild or moderate AS at a second tertiary hospital in Spain were used for validation. The timing of echocardiography surveillance based on the machine learning model was compared with European and American guideline recommendations for echocardiographic follow-up, and defined as premature (the patient did not develop severe AS within 6 months of the recommended follow-up interval), timely (scheduled follow-up was within 6 months of the diagnosis of severe AS), or untimely (the patient developed severe AS >6 months before the recommended follow-up).
Results:
In the development model, 1,124 patients (68.6%) had mild AS and 514 (31.4%) had moderate AS; follow-up was 3.6 ± 2.4 years; and AS progressed to severe in 581 (19.4%) echocardiograms, of which 39.8% were in year 1, 33% were in year 2, and 27.2% were in year 3 after the index echocardiogram. The derived model recommended follow-up at 1, 2, 3, or 4 years based on patient age and aortic valve Vmax, mean velocity, velocity time integral, left ventricular (LV) mass, mitral E-wave deceleration time, LV ejection fraction, LV stroke volume, mean LV outflow tract velocity, and LV end-diastolic volume. The validation cohort included 792 patients (51.7%) with mild AS and 741 (48.3%) with moderate AS; follow-up was 2.4 ± 1.7 years; and AS progressed to severe in 808 (27%) of echocardiograms, of which 49.1% were in year 1, 30.2% were in year 2, and 20.7% were in year 3 after the index echocardiogram. In the development population, the model discriminated the development of severe versus nonsevere AS at 1-, 2-, or 3-year intervals with a respective area under the receiver-operating characteristic curve (AUC-ROC) of 0.90, 0.92, and 0.92. In the validation population, the model had an AUC-ROC of 0.85, 0.85, and 0.85 for the 1-, 2-, or 3-year intervals. A simulated application of the model in the validation cohort suggested respective potential savings of 49% and 13% of unnecessary echocardiographic examinations per year compared with European and American guideline recommendations.
Conclusions:
The authors conclude that machine learning provides real-time, automated, personalized recommendations for timing the next echocardiographic follow-up examination among patients with mild or moderate AS; and that, compared with European and American guidelines, the model would reduce the number of echocardiographic examinations.
Perspective:
Taken at face value, this study suggests that, using machine learning, the rate of progression from mild or moderate AS to severe AS can be predicted based on patient age and nine variables from a single echocardiogram; and that fewer follow-up echocardiograms could be performed with an associated cost savings. To its credit, the study used separate derivation and validation cohorts. However, 1) AS severity was defined using only transaortic Vmax, 2) the study endpoint was echocardiograms with severe AS rather than individual patients progressing to severe AS, and 3) there was a surprising incidence of rapid progression to severe AS (e.g., in the validation cohort [52% with mild AS], 27% progressed to severe AS during a follow-up of 2.4 ± 1.7 years, 49% of which occurred during the first year). The methodology used to define AS and the apparently high incidence of rapid progression raise concerns about the reliability of the input data used for machine learning, and the use of echocardiograms as an endpoint rather than individual patients raises statistical concerns. Ultimately, confidence in predicting slower progression of AS needs to be balanced by the potentially catastrophic clinical outcomes for individual patients who do not follow a predicted course. Although this study raises the possibility that machine learning might be used to help predict AS progression, clinically useful follow-up recommendations need to include an assessment of patient outcomes.
Clinical Topics: Noninvasive Imaging, Valvular Heart Disease, Echocardiography/Ultrasound
Keywords: Aortic Valve Stenosis, Diagnostic Imaging, Disease Progression, Echocardiography, Heart Valve Diseases, Machine Learning, Stroke Volume
< Back to Listings
