Automated machine learning of echocardiographic strain enables identification of early myocardial changes in pre-symptomatic TTR carriers

This study demonstrates that an automated machine learning model applied to routine echocardiographic strain data can successfully identify subtle, pre-symptomatic myocardial abnormalities in individuals carrying the TTR V142I variant, offering a scalable strategy for the early detection of cardiac amyloidosis risk before clinical symptoms appear.

The Gist

The Big Picture: Finding the "Ghost" Before the House Collapses

Imagine a house (your heart) that is slowly being filled with heavy, sticky sand (amyloid protein) because of a specific genetic blueprint you were born with. This condition is called Transthyretin Amyloidosis (ATTR).

For a long time, doctors could only see the problem when the house was already caving in—when the walls were thick, the rooms were cramped, and the residents (the patient) were out of breath. By then, the damage was often too severe to fix easily.

However, a new study from Mount Sinai suggests we can spot the very first grains of sand entering the house, long before the walls even look different. They did this by using a special kind of "AI detective" to look at standard heart ultrasound videos.

---

The Characters in the Story

1. The Genetic "Smoking Gun" (TTR V142I):
   Think of the TTR gene as the instruction manual for building a protein. In some people (mostly African American and Hispanic/Latino populations), there is a typo in the manual called V142I. About 1 in 25 African Americans has this typo.

   • The Risk: If you have this typo, you have a 40–60% chance of eventually developing heart trouble from the sticky sand. But here's the catch: Having the typo doesn't mean you are sick yet. Many people carry it for decades without knowing.

2. The "Normal" Heart Scan:
   Doctors usually look at heart ultrasounds (echocardiograms) to check if the heart is pumping hard enough. It's like checking if a car engine is running. In the early stages of this disease, the engine looks like it's running perfectly fine. The car drives, but the sand is quietly piling up in the engine block.

3. The AI Detective (Machine Learning):
   This is where the study gets clever. The researchers didn't just ask, "Is the engine running?" They asked the AI to look at 200 tiny details in the ultrasound video that a human eye would miss.

   • The Analogy: Imagine a human looking at a crowd and saying, "Everyone looks happy." But an AI looking at the same crowd notices that 15 people are blinking slightly slower, 10 are leaning a tiny bit to the left, and the air temperature in the room is 0.5 degrees cooler. Individually, none of those things matter. But together, they signal that a storm is coming.

---

What Did They Do?

The researchers took two groups of people from a massive health database:

• Group A: People who have the genetic typo (TTR+) but feel perfectly healthy.
• Group B: People who do not have the typo (TTR-).

They fed thousands of heart ultrasound videos into a computer program.

• The Old Way: They first tried to find one single number (like "how hard the heart squeezes") that separated the two groups. Result: It failed. The numbers looked the same. The "sand" was too subtle to see with a ruler.
• The New Way (Machine Learning): They let the AI look at the pattern of movement across the whole heart. They used a technique called mRMR (which is like a smart filter that removes duplicate clues so the AI doesn't get confused).

The "Aha!" Moment

The AI found a hidden signature. Even though the heart looked normal to a human, the AI noticed a specific "dance" the heart muscle was doing that was slightly off.

• The Clues: The AI noticed that the bottom part of the heart (the base) was moving slightly differently than the tip (the apex). It also noticed tiny delays in how different layers of the heart muscle contracted.
• The Result: The AI could tell the difference between the "healthy" group and the "genetic risk" group with about 76–78% accuracy.
  • Think of it like this: If you asked a human to guess which of two identical-looking twins is the one with the genetic typo, they would be right 50% of the time (a coin flip). The AI was right about 3 out of 4 times.

Why Does This Matter?

1. Early Warning System: Currently, we wait until the patient has symptoms (shortness of breath, swelling) to treat them. By then, the "house" is already damaged. This tool could flag people years before they get sick.
2. Better Treatment: There are new medicines that stop the "sticky sand" from forming. These work best if you start taking them before the damage is done. This AI tool helps doctors decide who needs to start the medicine early.
3. Scalable: Ultrasound machines are everywhere. If we can train them to spot this "early dance," we can screen thousands of people without needing expensive, rare tests.

The Limitations (The "But...")

• It's not a crystal ball: The AI isn't perfect (it's not 100% right). It's a "risk score," not a diagnosis. It says, "Hey, this heart is moving in a weird pattern; let's check this person closer."
• Retrospective: They looked at old data. They need to test this on new patients in real-time to prove it actually prevents heart failure in the future.
• Not everyone gets sick: Even with the genetic typo, some people never get the disease. The AI is trying to find the ones who might get sick, which is a tricky needle-in-a-haystack problem.

The Takeaway

This paper is like discovering a new way to listen to a heart. For years, we only listened for the "loud noises" (symptoms). This study shows that if you listen to the whispers (tiny, complex patterns in the heart's movement) using a smart computer, you can hear the disease coming long before it screams.

It turns the "genetic lottery" into a manageable risk, giving doctors a chance to protect the heart before the damage is irreversible.

Technical Summary

1. Problem Statement

• Clinical Context: Transthyretin amyloidosis (ATTR), specifically the variant form (ATTRv) caused by the V142I mutation, is a progressive disease leading to cardiac amyloidosis (CA) and heart failure. The V142I variant is prevalent in African American (4%) and Hispanic/Latino (1%) populations, with a 40–60% lifetime risk of developing symptomatic disease.
• The Gap: Current diagnostic strategies are "phenotype-first," meaning patients are usually identified only after symptoms emerge or significant cardiac structural changes (like wall thickening) occur. By this stage, the disease is often advanced, and while new therapies exist, they are most effective when initiated early.
• The Challenge: There are no established guidelines or tools to risk-stratify pre-symptomatic TTR+ carriers. Previous studies showed that standard univariate echocardiographic metrics (e.g., Global Longitudinal Strain or GLS) fail to distinguish pre-symptomatic carriers from non-carriers because early amyloid deposition creates subtle, distributed abnormalities rather than isolated, gross defects.

2. Methodology

The study employed a "genotype-first" approach using a retrospective cohort from the BioMe biobank at Mount Sinai.

• Study Population:
  • Development Cohort: 49 TTR+ carriers (100% V142I) and 45 matched TTR- controls (matched by age, sex, and ancestry). All were asymptomatic with no prior diagnosis of heart failure or amyloidosis.
  • Validation Cohort: An external cohort of 115 participants (59 TTR+, 56 TTR-), including carriers of non-V142I variants.
• Data Acquisition:
  • Archived transthoracic echocardiograms were processed using Speckle-Tracking Echocardiography (STE) software (TomTec/Philips).
  • Approximately 200 candidate features were extracted, including:
    • Global and segmental strain values (basal, mid, apical).
    • Layer-specific strain (endocardial vs. mid-myocardial).
    • Regional gradients (e.g., apical-to-basal ratios).
    • Temporal metrics (time-to-peak strain, mechanical dyssynchrony).
    • Engineered features like relative apical sparing indices.
• Machine Learning Pipeline:
  • Feature Selection: To address the "curse of dimensionality" (200 features vs. ~90 samples) and mitigate overfitting, the authors used Minimal Redundancy Maximal Relevance (mRMR). This algorithm selected features that maximized relevance to the TTR+ status while minimizing redundancy among the features themselves.
  • Model Architecture: A Random Forest (RF) classifier was trained on the reduced feature set.
  • Validation Strategy: The model was evaluated using 5-fold cross-validation to minimize optimism bias and tested on the external validation cohort.
  • Performance Metric: Discriminatory power was assessed using the Area Under the Receiver Operating Characteristic Curve (AUC).

3. Key Contributions

• Shift from Univariate to Multivariate Analysis: The study demonstrates that while individual strain parameters (like GLS) cannot detect early disease, a multivariate pattern of subtle abnormalities can.
• mRMR Feature Selection: The application of mRMR successfully distilled ~200 high-dimensional strain features into a parsimonious set of 15 physiologically relevant features. This reduced overfitting and improved model interpretability.
• Identification of Subclinical Signatures: The model identified specific early markers that are not apparent in standard clinical reads, including:
  • Relative Apical Sparing: Preserved apical function relative to the base.
  • Inferolateral Strain Abnormalities: Specific reduction in strain in the basal inferolateral segments.
  • Layer-Specific Deformation: Differences between endocardial and mid-myocardial fiber function.
  • Mechanical Timing Heterogeneity: Gradients in the time-to-peak strain between the base and apex.
• Longitudinal Validation: The study showed that model scores increase over time in TTR+ carriers, particularly in those over age 55, suggesting the model captures the progression of subclinical amyloid deposition.

4. Results

• Univariate Analysis: Confirmed previous findings; no single conventional or strain metric (including GLS, RV strain, or LA strain) showed a statistically significant difference between TTR+ and TTR- groups.
• Model Performance (Development): The Random Forest model using the 15 mRMR-selected features achieved an AUC of 0.757.
• Model Performance (External Validation): The model maintained robust performance on the external cohort (n=115), achieving an AUC of 0.781 (95% CI: 0.688–0.869) with a high sensitivity of 0.983.
• Feature Importance: The model relied on complementary signals rather than a single dominant variable. Key drivers included apical-to-basal strain ratios, inferolateral strain reduction, and basal-apical timing gradients. Left Ventricular Ejection Fraction (LVEF) contributed only modestly, confirming that early disease does not present as global systolic dysfunction.
• Longitudinal Trends: In a subset of patients with serial echocardiograms, model scores increased over time (mean follow-up ~5 years). The rate of score increase was significantly higher in participants aged ≥55 compared to those <55.

5. Significance and Implications

• Early Detection Strategy: This work provides a scalable, non-invasive method to identify individuals at high risk for CA before the onset of symptoms or overt structural changes.
• Clinical Utility: The model serves as a risk stratification tool to prioritize TTR+ carriers for closer surveillance or confirmatory testing (e.g., cardiac MRI or biopsy), rather than acting as a definitive diagnostic tool.
• Therapeutic Timing: By detecting the disease in the pre-symptomatic phase, this approach could enable the earlier initiation of disease-modifying therapies (stabilizers or silencers), which are known to be most effective before significant myocardial damage occurs.
• Generalizability: The model's ability to detect non-V142I variants in the validation cohort suggests the underlying physiological patterns of early amyloidosis may be shared across different TTR mutations.
• Future Directions: The authors propose integrating this ML-based scoring system into health system workflows to automate the screening of genetically at-risk populations, moving toward a proactive, genotype-guided surveillance paradigm.

Limitations Noted: The study is retrospective with a modest sample size; the outcome was genotype status rather than confirmed clinical CA (due to incomplete penetrance and lack of routine diagnostic workups in the biobank); and echocardiograms were obtained for clinical reasons rather than systematic screening, introducing potential selection bias.