Deep Learning Echocardiographic Trajectories of Heart Failure With Preserved Ejection Fraction

Figures

• 

  Figure 1. Cohort Identification Workflow After retrospective heart failure with preserved ejection fraction with longitudinal echocardiographic identification. The cohort data were split into training and validation. We applied the following deep learningtrajectory clustering pipeline. 1) Feature selection and transformation, after selecting complete cases and imputing missing values with a self-attention model, data were scaled and temporal changes computed. 2) Encoder compression, change over time data was encoded into embeddings. 3) Dimensionality reduction and trajectory clustering, Dimensionality reduction with Uniform Manifold Approximation and Projection, and trajectory clustering was performed (see also the central illustration for a visual representation). Trajectory subtypes were identified in the training and validation sets, then, correlated with clinical characteristics and outcomes to assess clinical relevance and methodological robustness. ∗Complete cases defined as patients with LVIDd and interventricular septum thickness at end-diastole in at least 2 time points. BNP = brain natriuretic peptide; Echo = echocardiogram; HF = heart failure; ICD codes = International…

• 

  Figure 2. Longitudinal Phenomapping Echocardiogram Clusters The encoder dimensionality reduction figure illustrates dimensionality reduction utilizing Uniform Manifold Approximation and Projection as the encoder. The X, Y, and Z axes represent components generated by Uniform Manifold Approximation and Projection. Clusters identified through deep learning longitudinal clustering analysis are visually by colors. Seven echocardiographic parameters (interventricular septum thickness at end-diastole, LVIDd, PW, pulmonary artery systolic pressure, tricuspid annular plane systolic excursion, left ventricular ejection fraction, and left ventricular mass index.) were used in a deep learning longitudinal clustering approach. According to the performance scores, k = 3 was the optimal number based on performance metrics.

• 

  Figure 3. Echo-Trajectories Derived From Longitudinal Phenomapping (A to H) Multiple panels illustrate the Echo-Trajectories derived from deep learning longitudinal clustering and single echocardiogram parameters using penalized B-spline curves, with 95% CIs depicted on the y-axis, and time represented in years on the x-axis. Each Echo-Trajectory is depicted in a unique color, providing insight into the patterns of cardiac structure and function over time. IVSd = interventricular septum thickness at end-diastole; LVIDd = left ventricular internal dimension at end-diastole; LVIDs = left ventricular internal dimension at end systole; LVM Index = left ventricular mass index/height, g/m2.7; PASP= pulmonary artery systolic pressure; PB-Spline = penalized B-spline; PW = left ventricular posterior wall; TAPSE = tricuspid annular plane systolic excursion; other abbreviation as in Figure 1.

• 

  Figure 4. Multivariate Logistic Regression Models for Baseline Characteristics and Echo-Trajectories Forest plot illustrates the three Echo-Trajectories as target and baseline characteristics as covariates in multivariate logistic regression models. A stepwise selection with backward elimination approach was used to identify baseline characteristics associated with specific Echo-Trajectories. Covariates reached a retention threshold of P = <0.05. Afib = atrial fibrillation; CKD = chronic kidney disease; COPD = chronic obstructive pulmonary disease; HLD = hyperlipidemia; MAP = mean arterial pressure; other abbreviations as in Figures 1 and 3.

• 

  Figure 5. Kaplan-Meier Survival Curve Illustrates All-Cause Mortality Across Different Echo-Trajectories The analysis reveals differences in survival rates among the groups (P < 0.001). Specifically, Trajectory 2 and 3 exhibited poor survival outcomes (P < 0.001) compared to Trajectory 1. Adjustment for multiple group comparisons was performed using the Tukey-Kramer test.

• 

  Figure 6. Echo-Trajectories and Laboratory Biomarkers The panels illustrate the dynamic trajectories of clinical laboratory parameters using penalized B-spline curves, with 95% CIs depicted on the y-axis, and time represented in years on the x-axis. Longitudinal clinical laboratory markers are shown as follows: panel (A), BNP; panel (B), Cr; panel (C), GFR; panel (D), Na+; panel (E), K+; and panel (F), QTc. The echo-trajectories are depicted in a unique color, providing insight into the patterns of cardiac function over time and its effect on clinical markers. It is worth noting that the trajectories are echocardiogram-derived, independent of laboratory parameters. Laboratory values were only compared after the echo-trajectories were established. Cr = creatinine; eGFR = estimated glomerular filtration rate; K+= potassium; Na+= sodium; QTc = QT corrected for heart rate; other abbreviations as in Figures 1 and 3.

• 

  Central Illustration. Patients With Heart Failure With Preserved Ejection Fraction Undergoing Serial Echocardiography in Routine Clinical Care Unsupervised longitudinal phenomapping revealed distinct remodeling trajectories with associated clinical profiles and outcomes. EHR = electronic health record; HFpEF = heart failure with preserved ejection fraction; LV = left ventricle; pHTN = pulmonary hypertension; RV = right ventricle; other abbreviations as in Figures 1, 3, and 4.