Explainable machine learning for estimation of elevated left ventricular filling pressure: a multicenter validation

Guideline-recommended algorithms (GL-algorithm) often results in indeterminate left ventricular filling pressure (LVFP). Despite high accuracy, machine learning (ML) methods lack interpretability, which necessitates the development of explainable ML models for clinical use.

To develop an explainable ML model for estimating LVFP, providing patient-level interpretation using gold-standard right heart catheterization (RHC) data.

We retrospectively enrolled 956 patients who underwent echocardiography and RHC at three hospitals within a median of 3 days. Two extreme gradient boosting models were trained using data from two hospitals (n = 621) to estimate elevated pulmonary artery wedge pressure (PAWP ≥ 18 mmHg) as a surrogate for elevated LVFP. Model 1 used variables from GL-algorithm, while Model 2 used variables selected based on Shapley additive explanations (SHAP) values. Models' area under the receiver-operating characteristic curve (AUROC) for elevated LVFP were compared using external test data from the other hospital (n = 335).

Overall, 31.0% had elevated PAWP, and 42.7% were classified as indeterminate LVFP by GL-algorithm, whereas the ML models classified all patients. AUROCs of Model 1 (0.82, 95% CI 0.73-0.92) and Model 2 (0.83, 95% CI 0.75-0.91) in classifiable cases by GL-algorithm significantly outperformed that of GL-algorithm (0.72, 95% CI 0.60-0.83, p = 0.020 and 0.016, respectively), Model 2 performed equally well for indeterminate cases. SHAP force plots visualized each variable's contribution to the ML model's decision for each patient.

Explainable ML outperformed GL-algorithm in estimating LVFP, providing a user-friendly tool for clinicians with patient-level interpretability.