Abstract
Aims: All-cause mortality ranges between 33% and 42% for individuals with untreated moderate to severe aortic stenosis (AS). Transcatheter aortic valve replacement makes this a treatable condition, if identified early. Machine learning-based tools show great promise to predict cardiovascular outcomes.
Methods and results: We developed and validated a machine learning model for 3-year prediction of AS risk (ASrisk) using serum biomarkers and vital sign measurements. We then evaluated the tool’s capacity to identify diagnoses of AS sequelae, echocardiographic outcomes in individuals not diagnosed with AS, as well as enrichment and 3-year aortic valve area reduction in individuals with high ASrisk. Among 919 954 participants, 429 996 were from the Mount Sinai Data Warehouse (MSDW) [2179 (0.5%) AS cases] and 489 958 were from the UK Biobank [5066 (1%) AS cases]. Odds ratio (OR) of AS sequelae increased quantitatively with ascending deciles of ASrisk [OR 1.63 (95% CI 1.60-1.67) in MSDW]. Increasing ASrisk by 1 SD resulted in higher odds of echocardiographic findings in undiagnosed individuals [OR 1.88 (95% CI 1.71-2.06) for Doppler velocity index, OR 2.50 (95% CI 2.36-2.64) for aortic valve area, and OR 2.61 (95% CI 1.89-2.71) for mean gradient]. Three years after risk assessment, individuals with ASrisk > 0.95 show an 11-fold enrichment for AS diagnosis in both cohorts and an average reduction in aortic valve area of 0.42 cm2.
Conclusion: ASrisk can predict risk of AS 3 years ahead of diagnosis in the general population.