This single-center study includes 49 asymptomatic and 51 pre-surgery PMR patients recruited between 2006 and 2010 under NHLBI Specialized Centers of Clinically Oriented Research grant (5, 6, 16). Primary degenerative mitral valve prolapse characterized by echocardiographic evidence of thickened, redundant leaflets with excessive motion and prolapse. Patients were excluded for evidence of: (1) aortic valve > trace aortic regurgitation or mean gradient of >10 mmHg, (2) mitral stenosis (mean gradient > 5 mmHg, valve area < 1.5 cm²), (3) endocarditis, (4) iatrogenic MR (ergot, radiation induced), (5) hemodialysis, (6) pregnancy, (7) presence of coronary artery disease (stenosis > 50%), 8) positive exercise tolerance test with myocardial perfusion. The Institutional Review Boards of the University of Alabama at Birmingham and Auburn University approved the study protocol. All participants gave written informed consent.

All data from patients’ baseline and return visits were obtained prospectively and recorded in electronic health data records. Cardiac magnetic resonance (CMR) imaging with tissue tagging was performed in control volunteers who had no prior history of cardiovascular disease or medical illness, no history of smoking, and no cardiovascular medications. Asymptomatic PMR patients had Class I status, with moderate/severe PMR by color flow Echo/Doppler, LVEF > 60%, LV end-systolic dimensions (ESD) < 40 mm, leaflet thickening and prolapse, and normal maximal exercise myocardial perfusion imaging (17). At baseline, PMR patients (asymptomatic, n = 49 and pre-surgery, n = 51), plasma xanthine oxidase (XO) activity and carboxy-terminal propeptide of procollagen type I (PICP), a marker of type I collagen synthesis, and carboxy-terminal telopeptide of collagen type I (ICTP) levels, a marker of type I collagen degradation were measured. Post-surgical CMR was performed six months after the surgical procedure in all pre-surgery patients.

Table 1 lists all CMR-derived LV and left atrial volumes, strains and twist as previously described in our laboratory (5, 6, 16, 17).

Peripheral venous XO activity was measured by the rate of uric acid production in the presence of xanthine (75 μM) without NAD⁺ as described previously in our laboratory (16).

Baseline levels of PICP and ICTP were measured with commercially available immunoassays (Quidel Corporation, USA and Orion Diagnostic, Finland). Detection limits were 0.2 ng/ml for PICP and 0.3 ng/ml for ICTP as described previously in our lab (16).

Data in Tables are presented as number (% of total) or median (interquartile range). Statistical differences between two groups are tested via Mann-Whitney U test for continuous variables and chi-square test for categorical variables (Tables 1, 2 and Supplementary Table S1). Comparisons between 3 groups are tested by Kruskal-Wallis and the p-values adjusted by false discovery rate (FDR) for multiple testing, reported in Supplementary Table S2. Univariate and multivariate logistic regression are fitted to predict the probability of post-surgical LVEF < 50% using 4 pre-surgical parameters (LVEF, LVESD, LV sphericity index, and LV systolic circumferential strain rate (Supplementary Tables S3, S4).

This study employed a standard LASSO (15) logistic regression and three machine learning models: random forest (RF) (12), support vector machine (SVM) with Radial Basis Function (RBF) Kernel (13), and extreme gradient boosting (XGBoost) (14). The models are trained to predict a binary response variable with two levels: LVEF < 50% or >50% at 6 months after mitral valve surgery in pre-surgery PMR patients. The models were initially developed using 37 features including demographics, CMR parameters, and biomarkers for predictive modeling (Table 1 and Figure 2).

To avoid overfitting and to increase the reproducibility of the models, we attempted to include the most relevant features in predicting post-surgery LVEF < 50%. LASSO, which is a regression model penalized on l1 penalty, performs both feature selection and regularization by itself to enhance the interpretability and accuracy of the linear model. Therefore, LASSO does not require extra feature selection. However, for RF, SVM, and XGBoost models, the most relevant features are selected by recursive feature elimination algorithm (18, 19). This fits machine learning models with all features at the beginning and excludes the least important features based on their importance rank, resulting in a model with a reduced number of features (Figure 2). This elimination process is then iterated by updating the feature importance rank using the model from the previous iteration, thus updating the model with a fewer number of features. The recursive feature elimination stops when the highest area under the ROC curve (AUROC) is achieved, and the final feature sets is the optimal feature set for our final models (Figure 3A).

For RF and XGBoost, the feature importance is measured by a mean decrease in Gini index, which reflects the contribution of each feature to the purity of the nodes or leaves in the tree-based models. The greater the mean decrease in Gini index, the more important the feature is in the model. For SVM, the feature importance is measured using the AUROC value. To compare the feature importance returned by different models, we scale the importance scores to the same scale (0–1) for each model separately as follows:Impi,scaled=Impi−ImpminImpmax−Impmin,i=1,2,…,37where Imp is the raw importance score, Impmax is the maximum importance score among 37 features, and Impmin is the minimum importance score among 37 features. Figure 3C shows the top features selected by the four models.

Fifteen percent of pre-surgical PMR patients are randomly selected as an independent testing set (n = 8), and the remaining 85% are used to train the initial machine learning models (n = 43) using the initial 37 features (Table 1). During each iteration in recursive feature elimination, hyper-parameter tuning and cross-validation are performed to obtain the model with a reduced feature set and optimal hyper-parameters. For the LASSO and machine learning methods, the corresponding models with the optimal feature sets that result in the highest AUROC were selected as final models. Subsequently, the four final models are further evaluated and compared via cross-validation. To eliminate the potential sampling bias and to obtain a robust model performance, we repeated the splitting, training, and testing process 100 times (i.e., using a different testing patient set of n = 8 vs. the training set of n = 43), and record the averaged model performance across iterations in the results.

The LASSO and machine learning models are evaluated via six metrics: classification accuracy, area under the receiver operating characteristic curve (AUROC), area under Precision-Recall (PR) curve (AUPRC), sensitivity, specificity, precision, and F1 score in the 51 pre-surgical PMR patients who have an actual known outcome of LVEF < 50% at 6 months post-surgery (Figure 3D). Among the six metrics, AUROC was used to choose the final prediction model for post-surgical LVEF < 50%.Accuracy=TP+TNTP+TN+FP+FN Sensitivity=TPTP+FNSpecificity=TNTN+FPPrecision=TPTP+FPF1score=2TP2TP+FP+FNWhere TP, TN, FP, and FN = true positive, true negative, false positive, and false negative.

To interpret the final predictive model, we compute the Shapley Additive exPlanations (SHAP) (14) value, which is developed from the Shapley value in cooperative game theory. The Shapley value in game theory quantifies the contribution that each player brings to the game. Similarly, the SHAP value quantifies the contribution of each feature to the model prediction. In our case, one game is one patient, and the players are the optimal feature sets in the final predictive model. The SHAP value of a particular feature is computed by weighting the marginal contributions of the feature. For pre-surgery PMR patient or asymptomatic PMR patient, we compute the SHAP value to evaluate the contribution of each feature to prediction of the post-surgical LVEF < 50%. Therefore, the SHAP value provides a local (i.e., patient-level) interpretation of the decision made by the final machine learning model. Moreover, by averaging the absolute SHAP values, we can obtain the overall feature importance, which provides an overall (i.e., group-level) interpretation of the machine learning model. In addition, the sign of the SHAP value implies the directional impact of each feature on the prediction (i.e., a positive SHAP implies positive impact on the probability of post-surgical LVEF < 50%).

There is a significantly higher incidence of episodic atrial fibrillation and medications, LV end-diastolic dimension (LVEDD), LVESD, and pulmonary artery systolic pressure (by Echo/Doppler) in pre-surgery vs. asymptomatic PMR. Forty seven per cent of pre-surgical patients are Class I, and all patients had normal renal function. Median pulmonary systolic pressure and median pulmonary artery wedge pressure are 38 and 16 mmHg, respectively, in pre-surgical patients (Supplementary Table S1).

LVEF does not differ among normal and both PMR groups. LV end-diastolic volume, LV stroke volume, LVEDD, LVESD, and regurgitant volume and XO activity are greater in pre-surgery vs. asymptomatic PMR. However, LVED mass/volume, Sphericity Index, and LVED 3-dimensional radius of curvature/wall thickness at mid LV do not differ in asymptomatic and pre-surgery PMR patients. Asymptomatic and pre-surgery patients have decreased LV systolic twist/volume slope (°/mL) vs. controls. The increase in plasma ICTP and decrease in the PICP are consistent with net collagen degradation in pre-surgery PMR patients. There is an increase in LA maximum and minimum volumes in asymptomatic PMR compared to normal and they are higher in pre-surgical PMR patients. However, only pre-surgical PMR patients have a decrease in total LA emptying fraction (Supplementary Table S2).

Among the 51 pre-surgical patients, 13 patients (25.49%) had a decrease in CMR LVEF < 50% at 6 months post-surgery. Patients with LVEF < 50% were more likely to have a median baseline LVEF < 58% by CMR and a greater incidence of atrial fibrillation. Those with a decrease in LVEF had a higher LV end-systolic volume index and lower LV sphericity index and LV mid systolic circumferential strain rate, prior to surgery. However, FDR-adjusted p-values indicated no significant differences between the two groups (Table 2).

Four predictive models: a standard linear model - LASSO, and three machine learning models: RF, SVM, and XGBoost were trained to predict a binary response variable with two levels: LVEF < 50% or >50% at 6 months after mitral valve surgery in the 51 pre-surgery PMR patients (Figure 2). The models were initially developed using 37 features including demographics, CMR parameters, and biomarkers for predictive modeling (Table 1). To reduce the complexity of the model, feature selection (Figure 2) identifies a subset of features that are relevant in predicting a post-surgical LVEF < 50% using LASSO for the linear model and recursive feature elimination for the machine learning models. Figure 3A shows the ROC and precision-recall curves (PRC) (Figure 3B) and the top important features selected by LASSO and the machine learning models (Figure 3C) and lists the model performance (Figure 3D) for all four models including: AUROC, AUPRC, accuracy, sensitivity, specificity, precision, and F1 score along with the 95% confidence interval generated via bootstrapping. Overall, the machine learning models outperform the LASSO regression model. Based on the top features, the RF model provided the highest AUROC and AUPRC and was chosen to re-train the 51 pre-surgical patients using the four most relevant features identified by RF, for predicting LVEF < 50%: baseline LVEF, LV Sphericity Index, LVESD, and LV Systolic circumferential strain rate to improve predictability of the final model (Figure 2).

The SHAP value interprets the contribution of the four important features of the RF model that predict post-surgery LVEF < 50% for both pre-surgery (Figure 4A) and asymptomatic PMR patients (Figure 4B). The directional impact of each feature is represented by the sign (negative or positive) of the SHAP value. A positive SHAP for each feature has a positive impact on the probability of post-surgery LVEF < 50%. A negative SHAP value has a negative impact on the probability of post-surgery LVEF < 50%. The larger the absolute value of SHAP, the greater the contribution the feature has to the model prediction.

Figure 4C shows the SHAP value of the four features in the RF model with a data-driven cutoff value (red vertical dashed line) indicating the directional impact on the likelihood of post-surgical LVEF < 50% in pre-surgery (top graphs) and asymptomatic (bottom graphs) PMR patients. Each circle represents one patient, and the color denotes positive (blue) or negative (red) impact of the feature on the probability (or likelihood) of post-surgery LVEF < 50%. The higher the absolute baseline CMR derived LVEF (>58%) and sphericity index (>1.3), the less likely for post-surgical LVEF < 50% (negative SHAP values). Mid LV systolic circumferential strain rate is a negative quantity and more negative values represent a greater circumferential shortening rate. Thus, <−0.63 1/s (−0.00063 1/ms) suggests that it is less likely for post-surgical LVEF < 50% (negative SHAP values). In comparison, the higher the LVESD (>48 mm), the more likely to have a post-surgical LVEF < 50% (positive SHAP value).

We address the important question of how many asymptomatic PMR patients would potentially be at risk for LVEF < 50% if they had gone to surgery. Therefore, we retrained the optimal RF model with the four selected features in all 51 pre-surgery PMR patients, and applied it to asymptomatic PMR patients (Figure 2). Random Forest predicted 14 out of 49 (28.57%) asymptomatic PMR patients would have LVEF < 50% post mitral valve surgery, had they gone to surgery.