Machine learning (ML) could improve the care of patients with valve diseases, as researchers demonstrated its ability to schedule an echocardiographic follow-up in patients with mild-to-moderate aortic stenosis (AS). The study investigators, who applied an ML model to predict whether these patients would develop severe valvular disease at 1, 2 or 3 years, found it was able to discriminate severe from non-severe AS development. Further external validation of the model’s ability in the patient cohort cut costs of unnecessary echocardiographic examinations per year by almost half (49%) when compared with European guideline recommendations and 13% compared with American guidelines. “To put numbers to these cost savings, if we applied our ML model to the European (741.4 million x 0.4%) and U.S. (327.2 million x 0.4%) AS groups, we estimate we could save 180,000 to 150,000 echocardiographic examinations per year,” the authors wrote The consequent cost saving in U.S. dollars would be $83,160,000 in Europe and $69,300,000 in the U.S. noted the study, published Monday online and in the June issue of JACC: Cardiovascular Imaging. Main findings Led by Antonio Sánchez-Puente, BSc, PhD, and P. Ignacio Dorado-Díaz, BE, PhD, from the University of Salamanca and the Carlos III Health Institute in Spain, study results revealed that in an internal validation, the model discriminated severe from non-severe AS development with an area under the receiver-operating characteristic curve (AUC-ROC) of 0.90, 0.92, and 0.92 for the 1-, 2- or 3-year interval, respectively. In the external application, the model showed an AUC-ROC of 0.85, 0.85, and 0.85, for the 1, 2, or 3- year interval. “Our ML system adds value to the current Level of Evidence: C guideline recommendations because it is able to replicate the results,” the team said, “and, in addition, it enables the possibility of tailoring the classification thresholds to schedule follow-up recommendations depending on the resource availability of each health care system.” The researchers also mentioned that current artificial intelligence models could provide information on Level of Evidence: C recommendations, which still represent 41.5% of the recommendations in major cardiovascular society guidelines with no meaningful improvement along the last 10 years. However, they also recognized that it was necessary to integrate future ML analysis with other recommendation grades. Study methodology The research team began the ML model generation by including demographic and echocardiographic patient data from 4,633 echocardiograms from 1,638 consecutive patients. These patients initially had mild-to-moderate AS and at least two periodic imaging assessments. Doppler echocardiographic transaortic peak velocity was used for defining AS severity, which was graded as mild from 2.0 to 2.9 m/s, moderate from 3 to 3.9 m/s, or severe when peak velocity was ≥4 m/s. Visit intervals were classed as 1 year if the next echocardiography was done in the 6 to 18-month interval, as 2 years in the 18- to 30-month interval, and as 3 years in the 30- to 42-month interval. Next, the team trained the ML model using feature selection, classification algorithm training, and hyperparameter tuning resulting in a final set of 10 variables selected. The variable set included peak aortic jet velocity, mean aortic velocity, aortic velocity time integral, patient age, left ventricular mass and slope of deceleration of the mitral E wave. Other variables taken into consideration included left ventricular ejection fraction, left ventricular stroke volume, mean left ventricular outflow tract velocity and left ventricular end-diastolic volume. XGBoost classifier algorithm selected To predict AS echocardiographic follow-up, the team used an XGBoost (open source software) classifier algorithm, chosen because of its proven performance and ability to update these models in the future. For the internal model validation, the individual ML classifiers were evaluated with a 10-fold cross-validation scheme with 10 repetitions. Predictions for severe AS development at 1, 2 and 3 years for the test set were then compared with the ground-truth labels. External model validation was carried out with an external cohort of 89,802 patients and 180,381 unselected consecutive echocardiograms from screenings between January 2011 to July 2019. Outcome data of the composite endpoint of aortic valve replacement (either surgical or percutaneous) or any-cause mortality were recorded from the electronic medical records for this cohort. Echocardiographic follow-ups scheduled by each system were labeled as premature if the patient did not develop severe AS within 6 months of the predicted interval. The follow-up was labelled as timely if the scheduled follow-up was within 6 months of the diagnosis of severe AS and untimely if the patient developed severe AS more than 6 months before follow-up. Personalized follow-up plans are the goal In an editorial comment, Attila Kovács, MD, PhD, and Márton Tokodi, MD, PhD, both from Semmelweis University in Budapest, Hungary, emphasized the desirability of personalized follow-up plans. This approach would be tailored to the predicted course of the disease instead of the rigid follow-up schemes enshrined in the current guidelines, they added. “Nevertheless, these tasks would require more advanced algorithms and massive databases,” they acknowledged. “Of note, the models proposed in this study were not specifically designed to analyze a series of studies of a given patient.” In providing further study feedback, the editorialists also highlighted certain ML model parameters that are related to the envelope of the transaortic Doppler interrogation by SHAP (SHapley Additive exPlanations). Adding that Doppler-based measurements are prone to errors, Kovács and Tokodi commented that it would have been useful to investigate whether simple logistic regression models incorporating only a few Doppler or structural parameters could achieve similar performance. The editorialists concluded by asking whether these models can be directly applicable to patients from other parts of the world and whether a similar approach could help optimize AS intervention and pinpoint patients who would benefit from early aortic valve replacement. “Patients with moderate AS experience surprisingly high rates of adverse outcomes, and along with advanced cardiovascular imaging, ML techniques should be inducted in the identification of high-risk patients,” they added. Sources: Sánchez-Puente A, Dorado-Díaz PI, Sampedro-Gómez J, et al. Machine Learning to Optimize the Echocardiographic Follow-Up of Aortic Stenosis. JACC Cardiovasc Imaging. 2023;16:733-744. Kovács A, Tokodi M. Refining Echocardiographic Surveillance of Aortic Stenosis Using Machine Learning: Toward Personalized and Sustainable Follow-Up Schemes. JACC Cardiovasc Imaging. 2023;16:745-748. Image Credit: ckybe – stock.adobe.com