An Interpretable AI Tool for SAVR vs TAVR in Low to Intermediate Risk Patients with Severe Aortic Stenosis

Background. Treatment selection for low to intermediate risk patients with severe aortic stenosis between surgical (SAVR) and transcatheter (TAVR) aortic valve replacement remains variable in clinical practice, driven by patient heterogeneity and institutional preferences. While existing models predict postprocedural risk, there is a lack of interpretable, individualized treatment recommendations that directly optimize long-term outcomes. Methods. We introduce an interpretable prescriptive framework that integrates prognostic matching, counterfactual outcome modeling, and an Optimal Policy Tree (OPT) to recommend the treatment minimizing expected 5-year mortality. Using data from Hartford Hospital and St. Vincent’s Hospital, we emulate randomization via prognostic matching and sample weighting and estimate counterfactual mortality under both SAVR and TAVR. The policy model, trained on these counterfactual predictions, partitions patients into clinically coherent subgroups and prescribes the treatment associated with lower estimated risk. Findings. If the OPT prescriptions are applied, counterfactual evaluation showed an estimated reduction in 5-year mortality of 20.3% in Hartford and 13.8% in St. Vincent’s relative to real-life prescriptions, showing promising generalizability to unseen data from a different institution. The learned decision boundaries aligned with real-world outcomes and clinical observations. Interpretation. Our interpretable prescriptive framework is, to the best of our knowledge, the first to provide transparent, data-driven recommendations for TAVR versus SAVR that improve estimated long-term outcomes both in an internal and external cohort, while remaining clinically grounded and contributing toward a more systematic and evidence-based approach to precision medicine in structural heart disease.

💡 Research Summary

This study presents an interpretable artificial intelligence (AI) framework designed to provide personalized treatment recommendations between Surgical Aortic Valve Replacement (SAVR) and Transcatheter Aortic Valve Replacement (TAVR) for patients with severe aortic stenosis at low to intermediate surgical risk.

The core challenge addressed is the lack of data-driven, prescriptive tools that directly recommend which treatment would optimize long-term outcomes for an individual patient, moving beyond predictive risk models for single procedures. To tackle the inherent confounding bias in observational data, the authors developed a multi-stage prescriptive pipeline.

First, patient data were integrated from two independent medical centers in Connecticut, USA: Hartford Hospital (training/internal validation cohort, n=236) and St. Vincent’s Hospital (external validation cohort, n=305). The cohort included patients with severe aortic stenosis (STS score <3.5) who underwent either SAVR or TAVR, featuring demographics, medical history, echocardiography, and computed tomography angiography (CTA) measurements. Mortality data was obtained via linkage to state death records.

The methodology consists of four key steps: 1) Prognostic Matching: To emulate randomization, patients were stratified by STS risk score and then matched 1:1 based on clinical covariates between the SAVR and TAVR groups, reducing selection bias. 2) Counterfactual Estimation: Separate Survival Random Forest models were trained for the SAVR and TAVR arms on the matched cohort. These models were used to estimate each patient’s counterfactual 5-year mortality risk under both treatment options. 3) Policy Learning: An Optimal Policy Tree (OPT) was trained using these counterfactual risk estimates. The OPT partitions patients into subgroups based on their features and prescribes the treatment (SAVR or TAVR) associated with the lower estimated mortality risk for each subgroup. 4) Sample Weighting: An iterative weighting scheme was applied to adjust for residual confounding, preventing the model from becoming overly optimistic for one treatment.

The results demonstrated the framework’s effectiveness. The learned policy tree achieved a 50% concordance with real-world clinical decisions in the Hartford cohort. More importantly, a counterfactual evaluation estimated that following the OPT’s recommendations could reduce 5-year mortality by 20.3% in the Hartford cohort and by 13.8% in the external St. Vincent’s cohort compared to the treatments actually received, indicating promising generalizability. The decision rules derived by the tree were clinically interpretable and aligned with established medical knowledge: TAVR was favored for frailer patients or those with low left ventricular ejection fraction (LVEF) and larger sinus of Valsalva diameter, while SAVR was preferred for less frail patients with preserved LVEF or those with a small aortic annulus.

The authors conclude that this interpretable prescriptive framework is the first of its kind to offer transparent, data-driven recommendations for the SAVR vs. TAVR decision, showing potential to improve estimated long-term outcomes. It exemplifies how integrating confounding adjustment methods with interpretable policy learning can leverage real-world data to advance precision medicine in structural heart disease. The study acknowledges limitations, including the restriction to patients with available CTA data and the need for prospective validation and incorporation of additional outcome measures like complications and valve durability.