A deep learning framework for standardized interpretation of multiparameter cardiac ultrasound and disease classification

Figures

• 

  Figure 1. Standardized quantitative measurement workflow and quality control system for cardiac ultrasound analysis The figure illustrates the standardized data acquisition environment designed to minimize measurement variability. (A) Real-time echocardiographic measurement following the standard operating protocol. (B) Standardized operator positioning setup designed to ensure stability and minimize acquisition error. (C) Ultrasound console calibration and parameter display interface used for system initialization. (D) Real-time quantitative measurement interface showing simultaneous data verification. (E) The clinical ultrasound system (Philips Affiniti 30) utilized for standardized image acquisition throughout the study. (F) Dual-screen configuration supporting real-time quality control and measurement validation during the examination.

• 

  Figure 2. Representative echocardiographic measurements and Doppler flow spectral analysis The figure displays a composite of standard 2D and Doppler echocardiographic assessments used for the structural and functional evaluation of the heart. These representative images illustrate the quantitative inputs processed by the DeepCard system. (A) Tissue doppler imaging (TDI) of the medial mitral annulus. The spectral waveform displays systolic (s') and diastolic (e', a') velocities, with the e' wave indicating normal diastolic relaxation properties. (B) M-mode measurement of the left ventricular end-diastolic dimension (LVEDD) obtained from the parasternal long axis view. Calipers are placed perpendicular to the long axis of the ventricle to ensure accurate quantification of chamber size. (C) Continuous wave (CW) Doppler spectrum across the aortic valve. The dense spectral envelope below the baseline during diastole indicates mild aortic regurgitation, characterized by a slow deceleration rate. (D) Pulsed wave (PW) Doppler of mitral valve inflow recorded at the leaflet tips. The distinct E-wave (early diastolic filling) and A-wave (atrial contraction) pattern demonstrates a normal E/A ratio,…

• 

  Figure 3. Scatterplot analysis of the correlation between physician clinical experience and diagnostic accuracy The graph depicts the relationship between professional clinical experience and overall diagnostic accuracy (y axis) for the participating clinicians, stratified into three experience strata (<3 years, 3–10 years, and >10 years). Each data point represents a physician’s diagnostic accuracy; the dashed line indicates the mean-accuracy trend and the shaded region its 95% confidence interval. Diagnostic accuracy was positively correlated with clinical experience (Pearson’s correlation: r = 0.68, R2 = 0.46, two-sided p < 0.001, N = 656), confirming that diagnostic proficiency generally improves with accumulated experience. The substantial dispersion of data points, particularly at the early-to-mid career stages, highlights the inter-observer variability inherent in manual echocardiographic interpretation, underscoring the clinical need for standardized AI-assisted diagnostic tools.

• 

  Figure 4. Comprehensive framework of the DeepCard system: from standardized data acquisition protocol to multi-task interpretation and reporting The diagram summarizes the full workflow integrating clinical acquisition infrastructure, protocol-driven measurement standardization, and the DeepCard multi-task learning core for AI-standardized interpretation of quantitative echocardiographic parameters. (A–C) Hardware configuration utilizing the Philips Affiniti 30 ultrasound system. These certified acquisition systems ensure high-resolution data collection and adherence to consistent imaging protocols. (D–F) Standardized acquisition and measurement protocol. The images depict the consistent positioning and scanning protocols maintained by clinicians to minimize inter-operator variability during echocardiographic data acquisition. (G) The core algorithmic architecture of DeepCard. The model processes a 39-parameter quantitative input vector through a multi-stage pipeline: (left) preprocessing and representation learning using 1D convolutional residual blocks (kernel size = 3, stride = 1) with ReLU activation, max-pooling, and dropout for robust feature encoding; the 39 features are ordered by…

• 

  Figure 5. Receiver operating characteristic (ROC) curves for multi-class disease severity grading The figure illustrates the diagnostic performance of the DeepCard system in distinguishing between five clinical severity levels across eight specific cardiac conditions. Each panel plots the true positive rate against the false positive rate. Line styles and severity classes: The curves correspond to specific severity grades as follows: green dashed line = mild (Mi); teal crossed line = mild-moderate (Mi-Mo); red solid line = moderate (Mo); orange crossed line = moderate-severe (Mo-Se); and blue solid line = severe (Se). The area under the curve (AUC) values for each severity class are inset within the respective panels, demonstrating the model’s discriminative ability. (A) Mitral valve regurgitation. (B) Tricuspid valve regurgitation. (C) Aortic valve regurgitation. (D) Aortic valve stenosis. (E) Pulmonary valve regurgitation. (F) Left ventricular diastolic dysfunction. (G) Pulmonary artery hypertension. (H) Left ventricular systolic dysfunction.

• 

  Figure 6. Receiver operating characteristic (ROC) curves with 95% confidence intervals for binary cardiovascular classification tasks The figure evaluates the model’s diagnostic performance across nine binary classification tasks. Each pair of panels corresponds to a specific cardiac parameter, distinguishing between the presence of pathology and the normal state. Color coding: the blue curves (left : A, C, E, G, I, K, M, O, Q) represent the performance for detecting the presence (“Disease”) of the specified abnormality. The red curves (right: B, D, F, H, J, L, N, P, R) represent the performance for identifying the absence (“None” or normal state) of the condition. Error bars: Vertical error bars depict the 95% confidence intervals, indicating the stability and robustness of the model’s predictions across the validation dataset. AUC values are inset in each panel. (A–B) Pulmonary artery pressure (abnormal vs. normal). (C–D) Left atrial enlargement. (E–F) Right atrial enlargement. (G–H) Left ventricular enlargement. (I–J) Right ventricular enlargement. (K–L) Left ventricular hypertrophy. (M–N) Interventricular septal thickness. (O–P) Pericardial effusion. (Q–R) Regional wall motion…

• 

  Figure 7. Confusion matrices for nine binary cardiovascular classification tasks The figure visualizes the concordance between the model’s predictions and the ground truth labels across nine binary diagnostic tasks. Matrix interpretation: For each panel, the y axis represents the actual class (ground truth), and the x axis represents the predicted class. The color intensity (gradient from light pink to dark red) reflects the sample density, with darker red indicating a higher number of patient cases. Performance indication: the elements along the main diagonal (top-left to bottom-right) represent correct classifications (true negatives and true positives). Off-diagonal elements represent misclassifications (false positives and false negatives). The high concentration of darker colors along the diagonal across all panels demonstrates the model’s high accuracy and balanced performance. (A) Pericardial effusion (PE). (B) Regional wall motion abnormality (WMA). (C) Ventricular septal defect (VSD). (D) Left atrial enlargement (LAE). (E) Right atrial enlargement (RAE). (F) Left ventricular enlargement (LVE). (G) Right ventricular enlargement (RVE). (H) Left ventricular hypertrophy (LVH). (I)…

• 

  Figure 8. Confusion matrices for multi-class disease severity grading The figure displays the classification performance for four key hemodynamic tasks, comparing the model’s predicted severity grades against the ground truth. Matrix interpretation: the y axis represents the actual severity class, and the x axis represents the predicted severity class. The matrices cover five severity levels: mild, mild-moderate, moderate, moderate-severe, and severe. The color intensity (darker red) indicates a higher concentration of cases. Diagonal dominance: The distinct dark diagonal pattern across all panels signifies high classification accuracy. Notably, misclassifications (lighter blocks off the diagonal) are primarily concentrated in adjacent severity categories (e.g., misclassifying “Moderate” as “Mild-Moderate”), reflecting the model’s logical consistency in handling borderline cases rather than random error. (A) Mitral valve regurgitation (MR). (B) Tricuspid valve regurgitation (TR). (C) Pulmonary valve regurgitation (PR). (D) Pulmonary artery pressure (PAP).

• 

  Figure 9. Integrated visualization of feature distribution and diagnostic performance metrics across multi-task cardiac assessments Circos-style visualization of the model’s feature space and prediction confidence. The outer ring segments correspond to specific diagnostic tasks (e.g., LVSD, MR, TR). Inside each segment, the scatter distribution represents individual patient cases, where red dots (“upregulated”) indicate high-confidence positive predictions or high feature activation, and blue dots (“downregulated”) indicate negative predictions or low activation. The inner ring heatmap displays the aggregate Z score, reflecting the statistical significance of the diagnostic features for each category. Comprehensive performance metrics comparison. The chart combines a bar plot and line plots to display five key evaluation metrics for each cardiac condition. Blue horizontal bars: represent sensitivity (recall), mapped to the bottom x axis. Colored lines: represent specificity (dark red), accuracy (orange), precision (light blue), and F1 score (gray), mapped to the top x axis metric values. The chart highlights the model’s conservative diagnostic profile, where specificity (red line)…

• 

  Figure 10. Global interpretability analysis using SHAP feature attribution across 17 diagnostic tasks The heatmap matrix visualizes the relative importance of quantitative echocardiographic input parameters for each specific diagnostic output, revealing the model’s decision-making logic. Matrix layout: The y axis lists the 17 diagnostic tasks (including valvular, ventricular, and structural assessments), and the x axis represents the input echocardiographic measurements. Color coding: the color intensity represents the mean absolute SHAP value (feature importance). Darker or more intense colors indicate a higher contribution of a specific feature to a specific task, while lighter colors indicate minimal influence. Physiological coherence: the distinct clustering of high-importance features (dark blocks) for specific tasks demonstrates that DeepCard learns physiologically relevant associations—for example, prioritizing left atrial and mitral valve parameters for mitral regurgitation diagnosis—rather than relying on a generic set of global predictors.

• 

  Figure 11. Local interpretability analysis using SHAP summary plots for four representative diagnostic tasks The figure displays the “beeswarm” plots for the top contributing features in four key tasks. Plot interpretation: The features are ranked on the y axis by global importance (mean absolute SHAP value). The x axis represents the SHAP value, indicating the impact on the model output: positive values (right of the centerline) increase the predicted probability of the pathology, while negative values (left) decrease it. Each dot represents a single patient instance. Color coding denotes the actual value of the feature: red = high feature value; blue = low feature value. (A) Pulmonary artery pressure (PAP). The model prioritizes right-sided and pulmonary vascular dimensions (e.g., wider IVC or pulmonary artery diameters shown in red) as positive predictors for elevated PAP. (B) Left ventricular diastolic dysfunction (LVDD). High importance is assigned to left atrial size indices and Doppler filling parameters, consistent with guidelines for grading diastolic function. (C) Left ventricular systolic dysfunction (LVSD). The prediction is driven primarily by global functional markers (e.g.,…

• 

  Figure 12. Statistical consistency analysis of disease severity scoring across physician specialties and patient demographics The figure presents violin plots overlaid with boxplots to visualize the distribution of ground-truth severity labels, assessing consistency among three groups of diagnosticians: electrophysiologists, interventional cardiologists, and heart failure specialists. Plot interpretation: the width of the violin shape represents the data density (kernel density estimation); the internal boxplot displays the median (thick horizontal line) and interquartile range; the red dot indicates the mean score. Data are shown as distributions of severity scores; boxplots indicate the median and interquartile range, and red dots indicate the mean. (A–D) Comparison of severity scoring standards among the three physician specialties across four disease categories: (A) ventricular dysfunction, (B) valvular heart disease, (C) other cardiac abnormalities, and (D) myocardial structural abnormalities. Group differences were assessed by Welch’s ANOVA (Welch’s F-test) with Games-Howell post-hoc pairwise comparisons; partial omega-squared (ωˆp2) is reported as the effect size. No significant…