Automated Phenotyping of Mitral Stenosis Using Deep Learning

This study introduces EchoNet-MS, an open-source deep learning framework that accurately and robustly automates the detection of mitral stenosis severity and its rheumatic etiology across multiple diverse echocardiographic cohorts, demonstrating strong potential as a clinical decision support tool.

The Gist

Imagine your heart is a busy house with four rooms, and between two of those rooms sits a special door called the mitral valve. This door is supposed to open wide to let blood flow through and then snap shut tight to keep it from leaking back.

The Problem: A Stuck Door
Sometimes, this door gets stiff, thick, or scarred. It doesn't open all the way. This condition is called Mitral Stenosis (or MS). It's like trying to drink a thick milkshake through a tiny, clogged straw. The heart has to work much harder to push blood through, which can lead to serious trouble like heart failure or blood clots.

Doctors use a special camera called an ultrasound (echocardiogram) to look at this door. They take videos from different angles to see how stuck it is. But here's the catch:

1. There are a lot of these videos.
2. Reading them is like trying to solve a complex puzzle while running a marathon; it takes a long time and can be tiring.
3. Sometimes, different doctors might look at the same video and disagree on how bad the clog is.
4. There are two main reasons the door gets stuck: Rheumatic fever (an old infection that scars the door) or aging/calcium buildup (like rust on an old gate). Knowing which one it is changes how the doctor treats the patient, but telling them apart is tricky.

The Solution: The AI Detective (EchoNet-MS)
The researchers in this paper built a super-smart computer program called EchoNet-MS. Think of it as a tireless, super-observant detective who has watched millions of heart videos.

Here is how it works, using a simple analogy:

1. The "Eyes" (The Camera Team)

Imagine you are trying to judge how stuck a door is. You wouldn't just look at it from one angle. You'd look from the side, from the front, and maybe even use a flashlight (Doppler) to see how the air (blood) is moving.

The AI does the same thing. It has six different "eyes" (neural networks) that look at the heart videos:

• Two eyes look at the door from the side (PLAX view).
• Two eyes look at the door from the front (A4C view).
• Two eyes use "flashlights" (Color Doppler) to see the speed of the blood flow.
• One eye even looks at a still picture of the blood flow speed.

2. The "Brain" (The Team Huddle)

Once each "eye" takes a look, they don't just guess on their own. They all meet in a "team huddle" (an ensemble model). They share their opinions: "I think it's pretty stuck," says one. "I see some calcium," says another. "The blood is moving really fast," says a third.

The AI combines all these opinions into one final, highly accurate verdict. It can tell you:

• How bad is the clog? (Mild, Moderate, or Severe?)
• Why is it stuck? Is it the old infection (Rheumatic) or just age/rust (Calcific)?

3. The Training School

To get this AI ready for the real world, the researchers didn't just show it a few pictures. They fed it a massive library of over 430,000 heart videos from thousands of patients across three different major hospitals in California. It's like the detective reading every single case file in the city before ever stepping out to solve a crime.

The Results: Why This Matters

When they tested this AI on new patients it had never seen before, it was incredibly accurate.

• The "Safety Net": The AI is amazing at saying, "This patient is definitely fine." It rarely misses a serious case. This is crucial because if a doctor misses a severe clog, the patient could get very sick.
• The "Speedster": It can do this work in seconds, freeing up human doctors to focus on treating the patient rather than staring at screens for hours.
• The "Universal Translator": It worked just as well at different hospitals, even though those hospitals used slightly different machines and had different types of patients. This means it's not just a "one-trick pony" for one specific hospital; it's a general tool that can work anywhere.

The Bottom Line

This paper introduces a new digital assistant for heart doctors. It acts like a second pair of eyes that never gets tired, never misses a detail, and can instantly tell the difference between a clogged door caused by an old infection versus one caused by aging.

While it's not ready to replace human doctors (doctors are still needed to make the final call and treat the patient), it acts like a high-tech safety net, ensuring that no one with a dangerous heart valve slips through the cracks. It's a big step toward making heart care faster, safer, and available to more people.

Technical Summary

1. Problem Statement

Mitral Stenosis (MS) is a significant valvular heart disease with substantial global health burdens, leading to complications such as atrial fibrillation, thromboembolism, and heart failure. Accurate diagnosis requires distinguishing between severity (mild, moderate, severe) and etiology (rheumatic vs. non-rheumatic/calcific), as these factors dictate different treatment strategies and procedural risks.

• Clinical Challenge: Comprehensive MS assessment via Transthoracic Echocardiography (TTE) is time-consuming, requires integrating multiple views (2D and Doppler), and suffers from inter-reader variability.
• AI Gap: While deep learning models exist for other valvular diseases (e.g., Aortic Stenosis, Mitral Regurgitation), robust, externally validated models specifically for automated MS severity grading and etiologic characterization have been limited. Existing foundation models often lack sufficient MS-specific training data.

2. Methodology

The authors developed EchoNet-MS, an open-source, end-to-end deep learning framework designed to automate MS phenotyping.

Data Sources & Cohorts

The study utilized a massive, multi-institutional dataset comprising 431,612 videos from 44,671 studies across three healthcare systems:

• Kaiser Permanente Northern California (KPNC): Used for training, internal validation, and temporal testing.
  • Derivation Set: 8,677 studies (7,576 patients).
  • Held-out Test: 1,623 studies.
  • Temporally Distinct Test: 34,371 studies (2025 data) to test temporal generalizability.
• External Validation:
  • Stanford Healthcare (SHC): 3,333 studies.
  • Cedars-Sinai Medical Center (CSMC): 72,909 studies.
• Ground Truth: Labels were extracted from finalized clinical echocardiography reports (severity and etiology) without additional core lab adjudication, reflecting real-world practice.

Model Architecture (Two-Stage Framework)

The framework integrates information from multiple echocardiographic views using a two-stage approach:

1. Stage 1: View-Specific Video/Image Classification

   • Video Models: Four separate Convolutional Neural Networks (CNNs) based on ResNet R(2+1)D architecture were trained to classify MS severity from four specific views:
     • Parasternal Long Axis (PLAX)
     • Apical Four Chamber (A4C)
     • PLAX Color Doppler
     • A4C Color Doppler
   • Image Model: A ResNet50 architecture was trained on static Continuous Wave (CW) Doppler images to predict mean mitral valve inflow gradients.
   • Etiology Model: A parallel set of video-based CNNs was trained specifically to classify Rheumatic vs. Non-Rheumatic MS.
   • Preprocessing: Videos were subsampled to 16 frames; images/videos were resized (112x112 for video, 224x224 for images); text was masked; and metadata was de-identified.

2. Stage 2: Ensemble Integration

   • Outputs from the six individual networks (4 video + 1 image for severity; 4 video for etiology) were fed into a HistGradientBoostingClassifier (from scikit-learn).
   • This ensemble approach aggregates predictions across views, handling missing views robustly and producing final probabilities for severity classes (None, Mild, Moderate, Severe) or etiology.

Training Details

• Framework: PyTorch 2.5.0 and PyTorch Lightning.
• Optimizer: Rectified Adam Schedule Free.
• Hardware: NVIDIA L40S GPUs.
• Loss Functions: Cross-entropy for classification; Mean Squared Error (MSE) for regression (gradient prediction).

3. Key Contributions

1. First Open-Source MS-Specific Framework: EchoNet-MS is the first open-source deep learning model specifically designed for automated MS severity grading and etiologic differentiation.
2. Massive Scale & Diversity: The model was trained on a significantly larger and more diverse dataset than previous foundation models (e.g., EchoPrime, PanEcho), which had limited MS-positive cases (901–2,097 studies) compared to EchoNet-MS's 5,239 MS studies in training.
3. Multi-View Integration: Unlike single-view models, EchoNet-MS synthesizes data from B-mode, Color Doppler, and CW Doppler across multiple anatomical views.
4. Etiologic Differentiation: The model uniquely addresses the distinction between rheumatic and non-rheumatic MS, a critical clinical nuance often missed by general phenotyping models.
5. Rigorous External Validation: The model was validated across four independent cohorts (including a temporally distinct 2025 cohort and two external institutions), demonstrating robust generalizability.

4. Results

The model demonstrated excellent performance across all cohorts:

• Severity Discrimination (Severe MS):
  • KPNC Held-out: AUC 0.937 (95% CI: 0.913–0.958).
  • KPNC Temporally Distinct: AUC 0.994 (95% CI: 0.986–0.999).
  • Stanford (SHC): AUC 0.991.
  • Cedars-Sinai (CSMC): AUC 0.973.
• Moderate or Severe MS Detection:
  • Consistently high AUCs ranging from 0.912 to 0.987 across cohorts.
• Negative Predictive Value (NPV):
  • Extremely high NPVs (0.899 – 0.999), indicating the model is highly reliable at ruling out clinically significant MS.
• Etiology Classification (Rheumatic vs. Non-Rheumatic):
  • Achieved AUCs ranging from 0.890 to 0.967 across cohorts.
  • High specificity (0.927–0.949) and NPV (0.984–0.995).
• Subgroup Analysis:
  • Performance remained robust across age groups, sexes, BMI categories, LVEF strata, and patients with concomitant valvular diseases (AS, MR, TR) or Atrial Fibrillation.
• Comparison with State-of-the-Art:
  • EchoNet-MS significantly outperformed existing foundation models (EchoPrime and PanEcho) on the held-out and temporally distinct cohorts (e.g., 0.937 vs. 0.846 for EchoPrime on held-out data, $p < 0.001$).
• Explainability:
  • Saliency maps confirmed the model focused on the mitral valve leaflets and corresponding Doppler signals, aligning with clinical reasoning.
  • Misclassifications were predominantly one-category errors (93–98%) and were more common in older patients.

5. Significance

• Clinical Decision Support: EchoNet-MS offers a high-throughput, reproducible tool to screen for clinically significant MS, potentially reducing missed diagnoses and facilitating timely referrals for intervention.
• Workflow Integration: The high NPV suggests the tool is particularly valuable for "ruling out" severe disease, allowing clinicians to prioritize high-risk cases.
• Generalizability: The successful validation across different healthcare systems, machine types (Philips EPIQ/Affiniti), and time periods suggests the model is robust to real-world variations in data acquisition.
• Open Science: By releasing the code and model weights (GitHub: https://github.com/echonet/MS), the authors enable further research, validation, and clinical integration by the broader community.
• Limitations & Future Work: The study is retrospective, relies on report-based ground truth (not core lab), and excludes patients with prosthetic valves. Future work requires prospective validation and integration into clinical workflows to assess impact on patient outcomes.