Prognostic value of artificial intelligence-derived echocardiographic measurements in transthyretin cardiomyopathy

This study demonstrates that fully automated, AI-derived echocardiographic measurements of left ventricular strain and right ventricular function provide independent and incremental prognostic value for risk stratification in transthyretin cardiomyopathy, performing comparably to human measurements and significantly enhancing existing biomarker-based staging systems.

The Gist

The Big Picture: A New "Weather Forecast" for a Heart Disease

Imagine your heart is a house. In a condition called Transthyretin Cardiomyopathy (ATTR-CM), a sticky, glue-like protein (amyloid) starts leaking into the walls of the heart, making them stiff and heavy. Over time, this "glue" makes the heart struggle to pump blood, leading to heart failure.

Doctors have been trying to figure out how fast this "glue" is building up and how likely the house is to collapse. Traditionally, they've used a blood test (biomarkers) to guess the severity. It's like looking at the humidity outside to guess if a storm is coming—it helps, but it's not the whole picture.

This study asks a new question: Can we use a special computer program (Artificial Intelligence) to look at ultrasound videos of the heart and predict the future better than the blood tests alone?

The Cast of Characters

1. The Patients: 347 people with this heart condition.
2. The Old Tool (NAC Staging): The current standard. It's like a traffic light system based on blood work.
   • Green: Low risk.
   • Yellow: Medium risk.
   • Red: High risk.
   • The Problem: Sometimes, two people have the same "Red" light, but one is fine for years while the other gets sick very quickly. The blood test can't tell the difference.
3. The New Tool (AI Echo): A computer software (called Us2.ai) that automatically watches ultrasound videos of the heart. It doesn't get tired, it doesn't need coffee, and it measures tiny movements of the heart muscle that human eyes might miss.

The Experiment: How They Tested It

The researchers took ultrasound videos of these patients' hearts and ran them through the AI software. The AI measured two specific things:

1. LV-GLS (Left Ventricle Strain): How much the main pumping chamber stretches and squeezes. Think of this as checking how flexible the rubber bands in a slingshot are. If they are stiff, the slingshot won't work well.
2. RV FAC (Right Ventricle Area Change): How well the smaller pumping chamber on the right side is working. Think of this as checking the backup generator.

They then created a new "Risk Score" based on these two numbers:

• Low Risk: Both the main pump and the backup generator are working well.
• Medium Risk: One of them is struggling.
• High Risk: Both are struggling.

The Results: The AI Saw What the Blood Test Missed

The study followed the patients for about 2.4 years to see who ended up in the hospital or passed away. Here is what they found:

• The AI was a crystal ball: The new AI-based score was incredibly good at predicting who would get sick.
• It added detail to the blurry picture: Even among patients who had the same blood test result (e.g., everyone was "High Risk" on the old scale), the AI could split them up.
  • Some "High Risk" blood-test patients were actually "Medium Risk" on the AI test and did okay.
  • Others were "High Risk" on the AI test and got very sick very fast.
• The Stakes: Patients the AI labeled as "High Risk" were 6 times more likely to have a bad outcome than those labeled "Low Risk."

The Analogy: Imagine two cars with the same "Check Engine" light on. The old blood test says, "Both cars are broken." The new AI test looks under the hood and says, "Car A has a loose belt (fixable), but Car B has a cracked engine block (disaster)." The AI helps doctors know which car needs immediate towing and which just needs a quick fix.

The "Human vs. Robot" Showdown

A major worry with AI is: "Is the robot actually good, or is it just guessing?"

To test this, the researchers compared the AI's measurements against measurements taken by expert human doctors looking at the same videos.

• The Verdict: They were equally good.
• The AI didn't make more mistakes than the humans. In fact, the AI was consistent every single time, whereas humans might get tired or distracted. This proves the AI is a reliable partner, not a replacement.

Why Does This Matter? (The "So What?")

1. Better Treatment Choices: There are expensive, powerful drugs available now to stop the "glue" from building up. However, they are very costly. This new AI test helps doctors decide who needs the expensive drug right now and who can wait. It prevents giving heavy-duty medicine to people who aren't sick enough yet, and ensures it goes to the people who are in the most danger.
2. No More Guessing: It turns a vague "maybe" into a clear "likely."
3. Speed and Consistency: Since the AI does the measuring automatically, doctors can get these precise results faster, without needing a specialist to spend hours analyzing every single frame of the video.

The Bottom Line

This study shows that using AI to measure how the heart muscle stretches and squeezes gives doctors a superpower. It adds a new layer of clarity to the diagnosis of heart amyloidosis.

Instead of just looking at the smoke (blood tests), the AI lets us see the fire (the actual heart muscle function). This helps doctors treat the right patients at the right time, potentially saving lives and money. The "Robot Doctor" is ready to join the team!

Technical Summary

1. Problem Statement

Transthyretin cardiomyopathy (ATTR-CM) is a progressive, fatal disease requiring accurate risk stratification to guide treatment. Currently, the standard for staging ATTR-CM relies on the National Amyloidosis Centre (NAC) system, which is based on biomarkers (NT-proBNP and eGFR).

• The Gap: While echocardiographic parameters like Left Ventricular Global Longitudinal Strain (LV-GLS) have shown prognostic value beyond biomarkers, their integration into clinical staging is not standardized.
• The AI Challenge: Artificial Intelligence (AI) tools for automated echocardiography are increasingly available, but less than 2% of approved AI medical devices perform a validated prognostic function. It remains unknown whether AI-derived measurements retain the prognostic accuracy of human measurements or offer incremental value over existing biomarker staging systems.

2. Methodology

Study Design & Population:

• Type: Retrospective cohort study.
• Data Source: Two Swiss ATTR-CM registries (Zurich and Bern).
• Cohort: 347 patients (91% male, median age 78) with confirmed ATTR-CM (92% wild-type, 8% hereditary).
• Follow-up: Median 2.4 years. Primary endpoint: Composite of all-cause death or heart failure hospitalization.

Data Acquisition & Analysis:

• AI Processing: Baseline echocardiograms were analyzed using Us2.ai, a fully automated AI software, to extract measurements.
• Comparison: AI-derived measurements were compared against routine human-measured reports from board-certified cardiologists.
• Biomarker Staging: Patients were staged using the NAC system (Stage I, II, III) based on NT-proBNP and eGFR.

Statistical Approach:

• Variable Selection: Multivariable Cox regression was used to identify independent predictors. Parameters were selected to avoid collinearity (e.g., LV structure, systolic function, diastolic function, RV function).
• Staging System Development: A systematic screening of 15 parameters combined with LV-GLS identified the optimal two-parameter model.
• Validation:
  • Risk Stratification: Kaplan-Meier analysis and Cox regression to assess hazard ratios (HR).
  • Incremental Value: Likelihood ratio chi-square tests and C-index improvements to determine if echo staging added value beyond NAC staging and age.
  • AI vs. Human: Receiver Operating Characteristic (ROC) curves and DeLong's test compared the 1-year predictive performance (AUC) of AI vs. human measurements.

3. Key Contributions

1. Validation of AI Prognostics: This is one of the first studies to validate that fully automated AI-derived echocardiographic measurements possess independent and incremental prognostic value in ATTR-CM, comparable to human measurements.
2. Novel Two-Parameter Staging System: The authors developed a simple, robust staging system combining LV-GLS and Right Ventricular Fractional Area Change (RV FAC).
   • Low Risk: Both parameters normal.
   • Intermediate Risk: One parameter abnormal.
   • High Risk: Both parameters abnormal.
3. Resolution of Heterogeneity: The study demonstrates that this echo staging system can stratify risk within established NAC biomarker stages, revealing significant prognostic heterogeneity that biomarkers alone miss.
4. Treatment Independence: The prognostic value of the AI-derived echo staging was confirmed to be independent of tafamidis treatment status.

4. Key Results

Study Outcomes:

• 141 patients (41%) experienced the composite endpoint (death or HF hospitalization).
• Independent Predictors: In multivariable analysis, AI-derived LV-GLS (HR 1.13 per 1% decrease) and RV FAC (HR 0.96 per 1% increase) were the only independent echocardiographic predictors of outcomes.

Performance of the Echo Staging System:

• Risk Stratification: The system successfully separated patients into distinct risk groups:
  • Intermediate Risk: 3-fold increased hazard (HR 3.0) compared to low risk.
  • High Risk: 6-fold increased hazard (HR 6.0) compared to low risk.
• Incremental Value: Adding echo staging to NAC staging and age significantly improved the model's predictive power (C-index increased from 0.730 to 0.772; $\chi^2$ improved from 53 to 80, $p<0.001$).
• Sub-staging within NAC: Within NAC Stage II, annualized event rates varied drastically based on echo risk (2% for low echo risk vs. 34% for high echo risk).

AI vs. Human Performance:

• AI-derived measurements showed comparable prognostic performance to human measurements for 1-year event prediction.
• AUC Comparison:
  • LV-GLS: AI (0.64) vs. Human (0.62), $p=0.65$.
  • LVEF: AI (0.66) vs. Human (0.69), $p=0.40$.
• No statistically significant differences were found across any tested parameters.

5. Significance and Clinical Implications

• Refining Risk Stratification: The study provides evidence that integrating AI-driven echo parameters (LV-GLS and RV FAC) into routine practice can refine risk assessment beyond current biomarker limits. This is crucial for identifying high-risk patients who may need aggressive therapy or closer monitoring.
• Scalability: Since AI software (Us2.ai) can automate these measurements without human intervention, this approach offers a scalable solution for consistent, reproducible risk stratification in clinical settings.
• Therapeutic Guidance: In an era of expanding disease-modifying therapies (e.g., tafamidis, gene silencers), accurate baseline risk stratification is essential for:
  • Identifying patients most likely to benefit from costly therapies.
  • Assessing treatment futility in advanced disease.
• Future Outlook: The findings support the integration of automated AI echo analysis into standard clinical workflows for ATTR-CM, though prospective validation in independent cohorts is recommended.

Conclusion: AI-derived echocardiographic measurements are not only comparable to human measurements but also provide critical, independent prognostic information that enhances existing biomarker-based staging systems for ATTR-CM.