Automated Classification of Mitral and Tricuspid Regurgitation With Explainability and Real‐World Practice Experience

Abstract

Background

Mitral regurgitation and tricuspid regurgitation frequently coexist and are evaluated using overlapping echocardiographic views. Although artificial intelligence–based approaches have shown promise, current existing models lack explainability and physiologic constraints, limiting their reliability and adoption in real‐world echocardiographic workflows.

Methods

We developed an end‐to‐end, explainable, flow‐aware deep learning pipeline that automatically selects relevant echocardiographic views from routine Digital Imaging and Communications in Medicine files, performs valve morphology assessment via leaflet pose estimation, incorporates systolic phase awareness for Doppler interpretation, and jointly grades mitral regurgitation and tricuspid regurgitation severity from transthoracic echocardiography. The model was trained and internally tested using 5086 outpatient studies from a tertiary center and externally validated in independent cohorts from 2 additional institutions. Model performance was evaluated using area under the receiver operating characteristic curve, sensitivity, specificity, positive predictive value, and negative predictive value, with subgroup analyses across key clinical strata.

Results

In the internal test set, the model demonstrated excellent discrimination for clinically significant (moderate or greater) mitral regurgitation and tricuspid regurgitation (area under the receiver operating characteristic curve >0.980), with high negative predictive values (>0.970). Performance remained robust across age, sex, and valvular comorbidities, with modest attenuation observed in patients with atrial fibrillation or reduced left ventricular ejection fraction. In the external cohort, areas under the receiver operating characteristic curve were modestly lower, particularly for severe regurgitation; however, sensitivity and negative predictive value remained high.

Conclusions

This explainable, unified artificial intelligence framework enables accurate and efficient screening of mitral regurgitation and tricuspid regurgitation in routine practice. By maintaining strong rule‐out capability and robust external performance, the proposed system supports automated triage and prioritization of echocardiographic studies in both specialist and noncardiology settings.

Nonstandard Abbreviations and Acronyms

A2C

apical 2‐chamber view

A3C

apical 3‐chamber view

A4C

apical 4‐chamber view

DELINEATE‐MR

Deep Learning for Echo Analysis, Tracking, and Evaluation of Mitral Regurgitation

KGH

Kuo General Hospital

MR

mitral regurgitation

NPV

negative predictive value

PPV

positive predictive value

PR

pulmonary regurgitation

RVI

right ventricular inflow

SAX

parasternal short‐axis

TNH

Tainan Hospital

TR

tricuspid regurgitation

VHD

valvular heart disease

Clinical Perspective.

What Is New?

• We developed a unified, flow‐aware artificial intelligence system that integrates view selection, valve morphology assessment, anatomic restriction, and systolic gating to automatically localize regurgitant jets and accurately grade mitral and tricuspid regurgitation from routine Doppler echocardiography.

• By restricting Doppler analysis to atrial regions during systole, the model substantially reduces misclassification of physiologic inflow or outflow and improves diagnostic performance, particularly for clinically significant regurgitation.

What Are the Clinical Implications?

• This explainable, end‐to‐end artificial intelligence pipeline enables rapid and reliable screening of mitral and tricuspid regurgitation in real‐world practice, allowing noncardiology clinicians to safely exclude significant disease while prioritizing high‐risk cases for specialist review.

Valvular heart disease affects ≈2.1% of the general population and up to 13.3% of individuals aged >75 years. ¹ However, community‐based echocardiographic screening identified a prevalence of only 1.8%, indicating that many cases remain undiagnosed. ² For patients with severe mitral regurgitation (MR) or tricuspid regurgitation (TR), timely recognition and intervention are essential for improving outcomes. ³ , ⁴ , ⁵ However, echocardiographic grading of regurgitation remains highly operator dependent and time consuming.

Recent advances in computer vision have enabled precision phenotyping of cardiac structure and function in echocardiography. Artificial intelligence (AI) models have been shown to accurately estimate regional wall motion abnormalities, predict imminent in‐hospital cardiac arrest, or generate realistic synthetic cardiac ultrasound images. ⁶ , ⁷ , ⁸ Building on this progress, several MR/TR‐oriented AI models have demonstrated strong performance in regurgitation assessment. For instance, the DELINEATE‐MR (Deep Learning for Echo Analysis, Tracking, and Evaluation of Mitral Regurgitation) study was designed to process complete transthoracic echocardiograms and apply a convolutional neural network that integrates spatiotemporal information across the cardiac cycle. This framework generates an MR severity determination, and its outputs have been shown to correlate with quantitative indices such as regurgitant volume. ⁹ Other approaches have focused on segmentation of MR jets on color Doppler images ¹⁰ ; however, frame‐level Doppler jet segmentation may be vulnerable to confounding by physiologic inflow that can resemble regurgitant flow. On the other hand, although earlier machine learning studies frequently relied on structured clinical variables for risk stratification in TR analysis, ¹¹ , ¹² , ¹³ recent deep learning work such as EchoNet‐TR has been proposed to evaluate TR directly from echocardiographic videos, challenging the notion that TR assessment is outside image‐based AI. ¹⁴

Since MR and TR frequently coexist and are evaluated using overlapping echocardiographic views, we developed an end‐to‐end pipeline that performs automated view selection, integrates valve morphology assessment via leaflet pose estimation, incorporates flow‐phase awareness, and jointly grades MR and TR severity to support efficient, reliable, and explainable clinical decision making. We hypothesized that incorporating explicit flow‐phase awareness, guiding the model to interpret Doppler signals in a physiologically consistent manner across the cardiac cycle, would reduce misclassification of normal inflow or outflow, improve diagnostic accuracy in borderline grades, and provide more clinically intuitive explanations than conventional processing.

METHODS

The data sets used and analyzed in the present study are available from the corresponding author upon reasonable request. The codes used in the present study are available at https://github.com/medkd/Valve‐regurgitation.

This study was conducted in accordance with the ethical principles of the Declaration of Helsinki. The study protocol was approved by the Institutional Review Board of National Cheng Kung University Hospital, Kuo General Hospital and Tainan Hospital, Ministry of Health and Welfare (approval reference: A‐ER‐109‐319). The requirement for informed consent was waived by the Institutional Review Board due to the retrospective nature of the study.

Data Set

Echocardiographic data (January 1, 2018, to December 31, 2018) were obtained from the cardiovascular databank of National Cheng Kung University Hospital, a tertiary hospital in Taiwan. Echocardiographic data were acquired in the left lateral decubitus position by using 1‐ to 5‐MHz probes (Vivid series, GE Healthcare; or EPIQ series, Philips Healthcare). The standard acquisitions included 3 apical views and parasternal long‐axis and short‐axis (SAX) views, archived in the Picture Archiving and Communication System at a resolution of 800×600 pixels.

The data set comprised 5086 outpatient transthoracic echocardiographic examinations performed at National Cheng Kung University Hospital in 2018, and patients with a history of prior valve intervention were excluded. To establish training and validation sets, each clip from these echocardiograms was processed using our previous view classifier, ¹⁵ which automatically labeled the apical 2‐chamber (A2C), apical 3‐chamber (A3C), apical 4‐chamber (A4C), SAX at the aortic valve level, and right ventricular inflow (RVI) views, along with their Doppler counterparts.

After automated classification, the initial MR cohort comprised 3382 echocardiograms and the initial TR cohort comprised 3567 echocardiograms. Cases with missing clinical characteristics or poor video quality were excluded. The final MR cohort included 2584 echocardiograms (1797 with no or mild MR, 528 with moderate MR, and 259 with severe MR), while the TR cohort included 2502 echocardiograms (1895 with no or mild TR, 436 with moderate TR, and 171 with severe TR). Each internal cohort was randomly divided into training, validation, and test sets in a 7:1:2 ratio (Figure 1).

Figure 1. Study cohorts.

Overview of the internal cohort—patients undergoing echocardiography at NCKUH in 2018—and external cohorts—patients undergoing echocardiography at KGH and TNH. The severity of regurgitation was graded and confirmed by experts. A2C indicates apical 2‐chamber view; A3C, apical 3‐chamber view; A4C, apical 4‐chamber view; ASE, American Society of Echocardiography; KGH, Kuo General Hospital; MR, mitral regurgitation; NCKUH, National Cheng Kung University Hospital; SAX, parasternal short‐axis view; TNH, Tainan Hospital; and TR, tricuspid regurgitation.

For temporal and external validation, we prospectively collected additional echocardiograms acquired at Kuo General Hospital (KGH) and Tainan Hospital (TNH), Ministry of Health and Welfare, between January 1, 2023, and December 31, 2023. These images were acquired using the same protocol.

Each study was independently reviewed by 2 board‐certified cardiologists following the guidelines of the American Society of Echocardiography who were not involved in the original clinical interpretation and were blinded to clinical assessments. ¹⁶ Regurgitation severity was determined using a guideline‐based hierarchy in which quantitative parameters (eg, effective regurgitant orifice area) were prioritized when available, with established qualitative indicators such as systolic pulmonary vein flow reversal. When these data were inconclusive or unobtainable, semiquantitative markers, including vena contracta width, were used for adjudication, with additional qualitative Doppler features considered supportive.

For cases in which the 2 reviewers assigned different severity categories, particularly in borderline ranges, the reference label followed the more severe classification. On the basis of institutional experience at our center, interobserver disagreement between experienced cardiologists occurs in ≈8% of cases. ¹⁷ Interpretations were recorded in a structured format, and regurgitation severity was classified as none/mild, moderate, or severe to improve the clinical applicability of our model ¹⁸ (Figure 1).

Workflow for Echocardiographic Analysis

To enable automated and clinically interpretable analysis of valvular disease, we developed an AI‐assisted echocardiographic workflow comprising 4 main components: view classification, cardiac structure and flow segmentation, systolic/diastolic event detection, and severity grading (Figure 2). Model development and reporting followed the Proposed Requirements for Cardiovascular Imaging–Related Multimodal AI Evaluation guideline and established standards for imaging‐based AI research. ¹⁹ In addition, this study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines.

Figure 2. Model structure.

After initial EfficientNet‐based view classification of the 20‐frame clip, the data were routed to either the MR or the TR branch. Subsequently, the event detector labeled each frame as systolic or diastolic. This was followed by chamber masking and regurgitant flow segmentation, which enabled precise localization and temporal mapping of jets; these quantitative features were used to drive the downstream regurgitation severity classifier. A2C indicates apical 2‐chamber view; A3C, apical 3‐chamber view; A4C, apical 4‐chamber view; CNN, convolutional neural network; LA, left atrium; LSTM, long short term memory network; LV, left ventricle; MR, mitral regurgitation; RA, right atrium; RV, right ventricle; SAX, parasternal short‐axis view; and TR, tricuspid regurgitation.

View Classification

Each echocardiographic clip was automatically assigned to its standard view to ensure appropriate downstream analysis and quality control. A classifier adapted from a pretrained EfficientNet‐B5 generated the view label. ¹⁵ , ²⁰ For MR analysis, A2C, A3C, and A4C and their color Doppler acquisitions were included; for TR analysis, A4C, RVI, and SAX views and their Doppler acquisitions were analyzed. ²¹ , ²² For views with multiple available clips, the clip exhibiting the highest weighted average mean frame‐level classifier probability was selected for downstream model input.

Cardiac Structure and Flow Segmentation

Valve leaflets, cardiac chambers, and regurgitation flow were delineated to support visualization. A compact network derived from the stacked‐hourglass architecture captured global geometry and fine structural details. ²³ Eight anatomic landmarks were annotated to facilitate valve localization, and segmentation masks of cardiac chambers and Doppler flow were subsequently applied for quantitative assessment of cardiac structure and regurgitant flow.

Systolic/Diastolic Event Detection

To differentiate systolic regurgitation jets from diastolic inflow signals, a temporal event detector was incorporated. This module identified systolic and diastolic frames from short clips, allowing the system to isolate clinically relevant frames for grading.

Integration and Severity Grading

Outputs from the systolic/diastolic event detector and segmentation modules were integrated to produce atrial‐restricted, systolic‐gated flow maps that highlight regurgitant regions. Severity of MR/TR was graded by an ensemble classifier that integrates cross‐stage–partial block and capsule network and aggregates spatiotemporal information across views while emphasizing computational efficiency and robustness to common acquisition variability. ²⁴ , ²⁵

Ablation Study: Input Configurations for MR/TR Severity Classification

To determine the most informative inputs for clinical grading, we conducted a predefined ablation study comparing 6 input configurations with the same internal validation data set: (1) raw B‐mode views only (A2C/A3C/A4C for MR; A4C/RVI/SAX for TR); (2) raw B‐mode views plus chamber masks; (3) raw B‐mode views plus valve leaflet localization; (4) color Doppler views only; (5) color Doppler views plus conventional flow segmentation (frame‐based, ungated); and (6) color Doppler views plus atrial‐restricted, systolic‐gated flow segmentation (phase‐aware, chamber‐restricted; our primary hypothesis).

Variables and Statistical Analysis

All models were trained for 100 epochs using the Adam optimizer (initial learning rate, 1×10⁻⁴), with adaptive learning rate reduction based on validation loss (decay factor, 1×10⁻¹; patience, 10 epochs; minimum learning rate, 1×10⁻⁶). Training was repeated 5 times, with performance reported as mean±SD. Model development and training were conducted using TensorFlow version 2.7.0, and statistical analyses were performed in Python version 3.8. Baseline characteristics—age; sex; body mass index; left ventricular ejection fraction (LVEF); and the prevalence of diabetes, hypertension, hyperlipidemia, stroke, coronary artery disease, chronic kidney disease, heart failure, atrial fibrillation (AF), and cancer—were compared across the training, validation, and test sets.

For the ablation study, F1 scores were computed on the internal validation set. Differences in model performance across input configurations were evaluated using the Kruskal–Wallis test, followed by post hoc pairwise comparisons with correction for multiple testing.

The input configuration with the highest F1 score for detecting clinically significant (moderate or greater) regurgitation was selected for severity grading. Confusion matrices, the area under the receiver operating characteristic curve (AUC) value and the calibration curve were used to evaluate severity classification model performance. In both internal and external test sets, we calculated the model's AUC values, F1 scores, recall (sensitivity) values, positive predictive values (PPVs), and negative predictive values (NPVs) for detecting moderate to severe and severe regurgitation.

In addition, subgroup analyses were performed to evaluate model performance across clinically relevant strata. Cases were stratified according to age (median, ≈45 years), sex, different LVEF, presence of AF at the time of image acquisition, ultrasound vendor, and coexisting valvular comorbidities, including MR, TR, aortic regurgitation, and pulmonary regurgitation.

Furthermore, 95% CI values were calculated through bootstrapping with 2000 samples. All statistical tests were 2‐tailed. A P value of <0.05 indicated significance.

To evaluate potential workflow benefits, we compared the average time required for manual MR and TR grading by experienced cardiologists with the total inference time of the proposed AI system. The end‐to‐end processing duration, from Digital Imaging and Communications in Medicine file loading to the generation of both severity classification and regurgitant jet visualization, was recorded for all test cases to quantify efficiency gains in clinical implementation.

RESULTS

Baseline Demographics

The baseline characteristics of the patients with MR or TR were similar between the internal training, validation, and test cohorts. The MR training, validation, internal test, and external test cohorts (n=1809, 259, 517, and 189, respectively) did not differ significantly in terms of age (46.75±15.23 versus 45.85±15.03 versus 45.37±14.53 years versus 47.31±12.42; P=0.213); sex distribution (men: 50.0% versus 52.9% versus 54.4% versus 49.2%; P=0.317); body mass index; the prevalence of major comorbidities such as diabetes, hypertension, hyperlipidemia, stroke, coronary artery disease, chronic kidney disease, heart failure, AF, or cancer (all P>0.05). On the other hand, the TR training, validation, internal test, and external test cohorts (n=1806, 224, 472, and 154, respectively) were well matched in terms of age (47.90±16.04 versus 46.21±15.31 versus 47.71±16.94 versus 47.57±14.91 years; P = 0.247), body mass index, LVEF, and other comorbidities. The distributions of MR and TR severity grades were consistent across the training, validation, and internal test sets, whereas the external test set exhibited a significantly different severity distribution (MR severity grading [none, mild, moderate, severe]: internal data set [37.0%, 32.6%, 20.4%, 10.0%], external data set [16.4%, 30.2%, 39.7%, 13.7%], P=<0.001; TR severity grading [none, mild, moderate, severe]: internal data set [51.2%, 24.3%, 17.7%, 6.8%], external data set [14.3%, 25.3%, 37.0%, 23.4%], P=<0.001; Table 1).

Table 1.

Baseline Demographic Characteristics of the Internal and External Cohort

	MR cohort					TR cohort
	Training	Validation	Internal Test	External Test	P value	Training	Validation	Internal Test	External Test	P value
	N=1808	N=259	N=517	N=189		N=1806	N=224	N=472	N=154
Age, y, mean±SD	46.75±15.23	45.85±15.03	45.37±14.53	47.31±12.42	0.213	47.90±16.04	46.21±15.31	47.71±16.94	47.57±14.91	0.247
Sex, male, n (%)	904 (50.0)	136 (52.9)	281 (54.4)	93 (49.20)	0.317	914 (50.6)	120 (53.6)	251 (53.1)	78 (49.67)	<0.001
Body mass index, mean±SD	26.18±6.13	25.64±5.56	25.48±6.65	24.73±4.39	0.139	26.05±5.12	25.61±5.52	24.97±7.31	26.56±7.35	0.182
LVEF, mean±SD	65.63±11.07	65.53±10.74	65.53±11.08	63.84±14.6	0.042	65.22±11.40	65.63±10.71	64.76±12.66	64.84±16.45	0.092
Diabetes, n (%)	308 (17.0)	39 (15.1)	83 (16.0)	22 (11.6)	0.257	332 (18.4)	34 (15.0)	79 (16.7)	23 (14.9)	0.447
Hyperlipidemia, n (%)	700 (38.7)	103 (39.9)	186 (35.9)	83 (43.9)	0.272	715 (39.6)	95 (42.3)	178 (37.8)	71 (46.1)	0.253
Hypertension, n (%)	693 (38.3)	105 (40.7)	192 (37.1)	53 (28.0)	0.033	717 (39.7)	92 (41.2)	163 (34.5)	66 (42.8)	0.128
Stroke, n (%)	80 (4.4)	8 (3.0)	25 (4.8)	4 (2.1)	0.313	87 (4.8)	8 (3.4)	25 (5.4)	4 (2.6)	0.458
Coronary artery disease, n (%)	260 (14.4)	41 (15.9)	70 (13.6)	20 (10.6)	0.420	278 (15.4)	37 (16.5)	78 (16.5)	23 (14.9)	0.908
Chronic kidney disease, n (%)	291 (16.1)	43 (16.7)	81 (15.6)	34 (18.0)	0.897	312 (17.3)	38 (16.9)	68 (14.4)	29 (18.8)	0.444
Heart failure, n (%)	360 (19.9)	51 (19.7)	105 (20.4)	34 (17.9)	0.921	392 (21.7)	45 (19.9)	105 (22.3)	26 (16.8)	0.494
AF, n (%)	175 (9.7)	19 (7.2)	53 (10.3)	21 (11.2)	0.519	193 (10.7)	19 (8.3)	57 (12.1)	15 (9.7)	0.524
Cancer, n (%)	284 (15.7)	40 (15.5)	81 (15.6)	25 (13.2)	0.846	287 (15.9)	35 (15.7)	89 (18.8)	27 (17.5)	0.452
MR severity, n (%)	660 (36.5)	98 (37.8)	197 (38.1)	31 (16.4)	<0.001
0 (none)	588 (32.5)	89 (34.4)	165 (31.9)	57 (30.2)
1 (mild)	374 (20.7)	50 (19.2)	104 (20.1)	75 (39.7)
2 (moderate)	186 (10.3)	22 (8.6)	51 (9.9)	26 (13.7)
3 (severe)
TR severity, n (%)						909 (50.3)	126 (56.1)	245 (51.9)	22 (14.3)	<0.001
0 (none)						437 (24.2)	46 (20.7)	126 (26.6)	39 (25.3)
1 (mild)						330 (18.3)	40 (17.9)	72 (15.2)	57 (37.0)
2 (moderate)						130 (7.2)	12 (5.3)	29 (6.1)	36 (23.4)
3 (severe)

AF indicates atrial fibrillation; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; and TR, tricuspid regurgitation.

View Classification and Systolic/Diastolic Event Detection

With EfficientNet‐B5, the view classifier achieved a mean accuracy of 94.8%, with the accuracy ranging from 99% for the A4C‐Doppler view to 85% for the SAX view. The systolic/diastolic event detector achieved a validation accuracy of 81.1%. These pretrained weights were subsequently applied to the regurgitant flow segmentation model (Table S1).

Cardiac Structure and Flow Segmentation

Valve leaflets, cardiac chambers, and regurgitant jets were delineated to facilitate visualization and quantitative analysis of valvular function. The stacked‐hourglass network achieved a mean squared error of 0.003 for valve localization (range, 0.001–0.008 across views; Figure 3A; Table S1).

Figure 3. Examples of echocardiographic leaflet estimation and segmentation for automated MR/TR detection.

A, The A4C view is presented as an example. The SHG‐based pose estimator localized the anterior and septal tricuspid valve leaflets by assigning 4 landmark points to each leaflet. B, In the same A4C view, replacing the terminal layer of the SHG architecture with a SoftMax layer enabled pixel‐wise segmentation of the cardiac chambers. C, Output of a single‐frame flow segmentation model (traditional method used in prior studies): misclassification of flow in the incorrect chamber. D, With the atrial‐restricted/systolic‐gated segmentation method, the proposed ensemble suppressed ventricular inflow and retained only true systolic regurgitant jets, markedly reducing the frame‐wise false‐positive rate across the clip. A4C indicates apical 4 chamber; ANT, anterior leaflet; LA, left atrium; LV, left ventricle; RA, right atrium; SEP, septal leaflet; SHG, stacked hourglass; and RV, right ventricle.

For chamber segmentation, the mean intersection over union was 0.852 (range, 0.817–0.904), while Doppler image segmentation yielded a mean intersection over union of 0.844 (range, 0.735–0.907). When applied to regurgitant flow segmentation, the model achieved intersection over union values of 0.973 for MR and 0.971 for TR (Figure 3B and 3C; Table S1).

Video Regurgitant Flow Segmentation

To capture temporal and spatial features, the systolic/diastolic event detector, chamber segmentation, and flow segmentation modules were combined into a unified workflow termed atrial‐restricted, systolic‐gated regurgitant flow segmentation. Compared with the conventional frame‐based segmentation model, which frequently misclassified ventricular inflow as regurgitant flow (Figure 3C; Video S1A and S1B), the proposed method reduced both false‐positive and false‐negative detections (Figure 3D; Video S1C and S1D).

Results of Ablation Analysis

Various input combinations were evaluated for classifying clinically significant MR and TR. The MR model obtained a mean F1 score (across all tested inputs) of 0.845; the highest score was 0.931 from A2C‐Doppler, A3C‐Doppler, and A4C‐Doppler views with corresponding atrial‐restricted, systolic‐gated flow segmentation, whereas the lowest score was 0.789 from A2C, A3C, and A4C raw B‐mode views with mitral valve leaflet localization. By contrast, the TR model yielded a mean F1 score of 0.812; the highest score was 0.916 from atrial‐restricted, systolic‐gated flow segmentation in the RVI‐Doppler, SAX‐Doppler, and A4C‐Doppler views, whereas the lowest score was 0.708 from inputs containing the RVI, SAX, and A4C raw B‐mode views (Table S2; Figure S1A and S1B). A Kruskal–Wallis test demonstrated significant differences in F1 score across configurations in MR and TR (P=<0.001). On the basis of these results, the 3 Doppler views with the proposed atrial‐restricted, systolic‐gated regurgitant flow segmentation method were selected for internal and external test set validation.

Model Performance in Grading the Severity of MR and TR

The severity grading framework exhibited robust and consistent performance for both valves. For MR, the model achieved internal AUC values of 0.981 (95% CI, 0.969–0.993) and 0.953 (95% CI, 0.945–0.961) for detecting moderate to severe and severe MR, respectively. The corresponding NPVs were 0.971 (95% CI, 0.966–0.979) and 0.982 (95% CI, 0.978–0.985), and the PPVs were 0.935 (95% CI, 0.919–0.947) and 0.660 (95% CI, 0.632–0.691), respectively. External validation revealed similar trends, with AUC values of 0.941 (95% CI, 0.930–0.953) and 0.969 (95% CI, 0.959–0.978), NPVs of 0.938 (95% CI, 0.928–0.950) and 0.986 (95% CI, 0.983–0.991), and PPVs of 0.832 (95% CI, 0.804–0.860) and 0.544 (95% CI, 0.517–0.576) for detecting moderate to severe and severe MR, respectively (Table 2; Figure 4A and 4C; Figure S2A).

Table 2.

Performance of the MR Severity Model on the Internal (NCKUH) and External (KGH/TNH) Test Sets

	Values (95% CI)
	Specificity	Recall	F1 score	AUC	PPV	NPV
Internal test set (NCKUH)
Moderate or greater MR	0.972	0.933	0.934	0.981	0.935	0.971
	(0.966–0.978)	(0.921–0.952)	(0.924–0.942)	(0.969–0.993)	(0.919–0.947)	(0.966–0.979)
Severe MR	0.954	0.838	0.740	0.953	0.660	0.982
	(0.947–0.961)	(0.632–0.695)	(0.720–0.760)	(0.945–0.961)	(0.632–0.691)	(0.978–0.985)
External test set (KGH/TNH)
Moderate or greater MR	0.780	0.955	0.889	0.941	0.832	0.938
	(0.735–0.822)	(0.950–0.963)	(0.871–0.909)	(0.930–0.953)	(0.804–0.860)	(0.928–0.950)
Severe MR	0.88	0.924	0.683	0.969	0.544	0.986
	(0.863–0.898)	(0.901–0.925)	(0.666–0.705)	(0.959–0.978)	(0.517–0.576)	(0.983–0.991)

AUC indicates area under the receiver operating characteristic curve; KGH, Kuo General Hospital; MR, mitral regurgitation; NCKUH, National Cheng Kung University; NPV, negative predictive value; PPV, positive predictive value; and TNH, Tainan Hospital.

Figure 4. Model performance for MR and TR severity classification.

A, ROC curves for detecting moderate to severe (dashed blue line) and severe (solid blue line) MR in the internal test set, along with corresponding curves from the external test set (dashed and solid orange lines). B, ROC curves for TR, plotted using the same format. C, Three‐class confusion matrix for MR (none/mild, moderate, or severe) in the external test set. D, Confusion matrix for TR in the external test set. Together, the 4 panels indicate that the model maintains high accuracy in differentiating between moderate to severe and severe regurgitation in both internal and external data sets for MR and TR. AUC indicates area under the receiver operating characteristic curve; MR, mitral regurgitation; ROC, receiver operating characteristic; and TR, tricuspid regurgitation.

For TR, the model achieved internal AUC values of 0.985 (95% CI, 0.976–0.993) and 0.968 (95% CI, 0.954–0.982) for detecting moderate to severe and severe TR, respectively. The corresponding NPVs were 0.992 (95% CI, 0.989–0.996) and 0.986 (95% CI, 0.981–0.990), respectively. External validation indicated strong model performance, with AUC values of 0.934 (95% CI, 0.921–0.947) and 0.917 (95% CI, 0.902–0.934), NPVs of 0.935 (95% CI, 0.901–0.969) and 0.968 (95% CI, 0.956–0.985), and PPVs of 0.797 (95% CI, 0.771–0.817) and 0.502 (95% CI, 0.480–0.523) for detecting moderate to severe and severe TR, respectively (Table 3; Figure 4B and 4D; Figure S2B).

Table 3.

Performance of the TR Severity Model on the Internal (NCKUH) and External (KGH/TNH) Test Sets

	Values (95% CI)
	Specificity	Recall	F1 score	AUC	PPV	NPV
Internal test set (NCKUH)
Moderate or greater TR	0.959	0.971	0.915	0.985	0.865	0.992
	(0.953–0.968)	(0.962–0.987)	(0.901–0.926)	(0.976–0.993)	(0.847–0.891)	(0.989–0.996)
Severe TR	0.987	0.778	0.787	0.968	0.797	0.986
	(0.984–0.993)	(0.706–0.859)	(0.734–0.836)	(0.954–0.982)	(0.760–0.881)	(0.981–0.990)
External test set (KGH/TNH)
Moderate or greater TR	0.623	0.971	0.875	0.934	0.797	0.935
	(0.567–0.666)	(0.959–0.986)	(0.855–0.892)	(0.921–0.947)	(0.771–0.817)	(0.901–0.969)
Severe TR	0.725	0.650	0.650	0.917	0.502	0.968
	(0.710–0.742)	(0.625–0.681)	(0.625–0.681)	(0.902–0.934)	(0.480–0.523)	(0.956–0.985)

AUC indicates area under the receiver operating characteristic curve; KGH, Kuo General Hospital; NCKUH, National Cheng Kung University; NPV, negative predictive value; PPV, positive predictive value; and TNH, Tainan Hospital; and TR, tricuspid regurgitation.

Calibration curves for MR and TR across the internal (National Cheng Kung University Hospital) and external (KGH/TNH) cohorts are also shown in Figure S3A and S3B. For MR, predicted and observed probabilities were closely aligned at higher predicted probabilities for both the moderate or greater and severe thresholds in both cohorts. In the low‐to‐intermediate prediction range, the calibration curves for both centers lay slightly below the identity line. The TR analysis demonstrated a comparable pattern.

The mean manual interpretation time for assessing MR and TR severity was 13.4 seconds and 14.3 seconds per study, respectively. In comparison, the proposed AI pipeline processed each Digital Imaging and Communications in Medicine view and produced both the severity classification and regurgitant jet overlay within 6.8 seconds for MR and 7.6 seconds for TR when implemented on an Nvidia RTX Titan GPU (Figure S4).

Subgroup Analysis

Subgroup analysis demonstrated that the model had robust performance across key clinical strata in internal and external test cohorts (Figures 5 and 6; Figure S5A and S5B) For MR evaluation in the external cohort, AUCs remained high in patients with preserved systolic function (LVEF >40%), reaching 0.948 for moderate or greater and 0.981 for severe MR. Performance was lower in LVEF ≤40% (0.931 for moderate or greater; 0.891 for severe). In AF subsets, overall discrimination remained strong; however, AUCs for moderate or greater and severe regurgitation were modestly lower than in patients without AF (moderate or greater MR, 0.918 versus 0.949; severe MR, 0.956 versus 0.972; Figure 5).

Figure 5. Forest plots showing subgroup performance of the MR detection model in the KGH/TNH cohort.

Model discrimination is assessed using the AUC with 95% CIs across predefined clinical subgroups. The left panel shows performance for detecting moderate or greater MR, and the right panel shows performance for detecting severe MR. Squares represent subgroup‐specific AUCs with corresponding 95% CIs, and the diamond indicates overall cohort performance. AF indicates atrial fibrillation; AR, atrial regurgitation; AUC, area under the receiver operating characteristic curve; KGH Kuo General Hospital; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; PR, pulmonary regurgitation; TNH, Tainan Hospital, Ministry of Health and Welfare; and TR, tricuspid regurgitation.

Figure 6. Forest plots illustrating subgroup performance of the TR detection model in the KGH/TNH cohort.

Discriminative performance is summarized using AUCs with 95% CIs across predefined clinical subgroups. The left panel presents performance for detecting ≥moderate TR, and the right panel presents performance for detecting severe TR. Squares denote subgroup‐specific AUC estimates with 95% CIs, and the diamond represents overall cohort performance. AF indicates atrial fibrillation; AR, atrial regurgitation; AUC, area under the receiver operating characteristic curve; KGH Kuo General Hospital; LVEF, left ventricular ejection fraction; MR, mitral regurgitation; PR, pulmonary regurgitation; TNH, Tainan Hospital, Ministry of Health and Welfare; and TR, tricuspid regurgitation.

Across age strata, model performance remained robust (age >45 years: AUC, 0.933 for moderate or greater MR and 0.958 for severe MR; age ≤45 years: AUC, 0.949 for moderate or greater MR and 0.980 for severe MR). Performance was similarly preserved across sex (men: AUC, 0.938 for moderate or greater MR and 0.974 for severe MR; women: AUC, 0.943 for moderate or greater MR and 0.966 for severe MR) and in the presence of concomitant aortic or pulmonary regurgitation. Vendor‐stratified analyses also demonstrated high AUCs (GE Vivid, 0.940 for moderate or greater MR and 0.970 for severe MR; Philips EPIQ, 0.943 for moderate or greater MR and 0.967 for severe MR; Figure 5). For TR evaluation in the external cohort, a similar overall trend was observed. However, within the LVEF subgroups, no obvious difference in model performance was noted between the moderate or greater TR and severe TR classifications (LVEF >40: AUC, 0.935 for moderate or greater TR and 0.912 for severe TR; LVEF ≤40: AUC, 0.931 for moderate or greater TR and 0.923 for severe TR; Figure 6).

DISCUSSION

Our end‐to‐end, flow‐aware deep learning model automatically selects relevant echocardiographic views from routine Digital Imaging and Communications in Medicine files, localizes 8 valve leaflet landmarks, and segments chambers and regurgitant jets with robust performance across multiple views. To further enhance physiologic fidelity, we incorporated an atrial‐restricted, systolic‐gated segmentation mechanism that more precisely delineates regurgitant flow. All processes were executed within a single unified software environment without the need for external echocardiography workstations or manual view selection. This design enables seamless integration into real‐world workflows, facilitates interpretation by noncardiology physicians, and substantially accelerates the grading process compared with manual evaluation.

In the internal test set, the MR and TR models both identified clinically significant (moderate to severe) disease with recall values of >0.930 and AUC values of >0.990 while maintaining NPVs above 0.970. When the classification threshold was restricted to severe disease only, both models retained strong rule‐out capability (NPV >0.980) for MR and TR, with AUC values of 0.965 and 0.988, respectively. Furthermore, subgroup analyses demonstrated robust performance for both MR and TR across most clinical strata, except in patients with reduced LVEF and AF.

Performance declined significantly in specific domains within the external cohort, most notably for severe TR (PPV, 0.502) and severe MR (PPV, 0.544). However, the models' high sensitivity was largely preserved (recall, 0.955 for moderate to severe MR and 0.971 for moderate to severe TR), and both models continued to miss few cases of truly significant regurgitation (NPV ≥0.935). Calibration analyses showed close agreement between predicted and observed probabilities at higher ranges for both MR and TR, whereas midrange probabilities in the external cohort tended to be overestimated, consistent with the moderate‐to‐severe misclassification patterns in the confusion matrices. Clinically, these profiles indicate that the model is better suited for rule‐out and triage, supporting its use as a screening and prioritization tool that reallocates cardiologist attention to cases requiring expert interpretation while preserving patient safety and reducing echocardiographic workload. ²⁶

Beyond specialist settings, the same triage principle may also hold great value in primary care and resource‐limited environments, where access to cardiology expertise is intermittent. In these settings, the high NPVs allow noncardiology physicians to confidently exclude clinically significant regurgitation at the point of care, while modest PPVs for severe disease suggest that AI‐flagged cases should prompt targeted specialist review rather than immediate intervention.

Several features differentiate the present study from previous echocardiographic AI studies. By generalizing across 26 views, our model eliminates the need for customized scan protocols or manual view selection that has been required in previous pipelines, ¹⁵ , ²⁰ enhancing the feasibility of bedside deployment. Training the atrial‐restricted, systolic‐gated network to recognize cardiac phase markedly mitigated the tendency of frame‐based systems to mislabel normal atrial inflow or ventricular outflow as regurgitation, ²⁷ , ²⁸ thereby reducing potential Doppler flow misclassifications. Furthermore, color Doppler clips augmented with regurgitant flow masks outperformed unannotated clips by highlighting jet momentum on a frame‐by‐frame basis, thereby directing model attention to pathophysiologically relevant regions and supporting explainability.

This study has several limitations. First, the training data were derived from a single tertiary center, and the overall sample size was smaller than in recently published large‐scale echocardiographic AI studies. ¹⁴ , ²⁹ These factors may limit generalizability and likely contributed to the modest performance decline observed in the external cohort. Second, all examinations originated from a Taiwanese population, and prior work has shown that Asian ventricular dimensions tend to be smaller than those of Western cohorts, ³⁰ which may influence Doppler jet appearance and affect cross‐population transferability. Third, morphological and quantitative characteristics of regurgitation were not systematically documented. Neither jet morphology (eg, eccentric or wall‐hugging patterns) nor pathogenetic subtypes (primary versus secondary disease) were consistently recorded, preventing subgroup analyses based on mechanism or jet behavior. Furthermore, although regurgitant jets may rarely occur during diastole, diastolic jets were not explicitly labeled, and a small number of such cases may have remained undetected. ³¹ To address these limitations, multicenter research is currently underway to prospectively collect larger and more diverse echocardiographic data sets incorporating hemodynamic parameters and pathogenetic annotations to further refine model generalizability and clinical applicability.

In conclusion, by integrating view selection, valve landmarking, chamber and flow segmentation, temporal‐phase detection, and severity grading into a unified pipeline, we advanced automated echocardiographic analysis beyond currently available single‐task systems. As for future work, we aim to explore extension of the model to a 5‐class system once appropriately annotated data sets become available to better capture clinical risk and treatment response, particularly in patients undergoing transcatheter tricuspid intervention. ³² In addition, incorporating structured murmur documentation and quantitative measures such as regurgitant volume or fraction in prospective studies may further contextualize the model's potential use in primary care and noncardiology settings. Furthermore, although leaflet pose estimation did not improve classification performance in the present study, it establishes an anatomic framework that may support future morphology‐driven analyses of valve disease as more detailed labels become available.

Sources of Funding

This study was supported by Taiwan's Ministry of Science and Technology (Grant Nos. 110‐2314‐B‐006‐090‐MY3 and 109‐2314‐B‐006‐068‐MY2), Ministry of Education and Headquarters of University Advancement at National Cheng Kung University (Grant No. D110‐G2512; under the Higher Education Sprout Project), and the Clinical Research Center at National Cheng Kung University Hospital (Grant No. NCKUH‐11002001/NCKUH‐11402057).

Disclosures

None.

Supporting information

Tables S1–S2

Figures S1–S5