Artificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis

This case report demonstrates that an externally validated artificial intelligence model applied to echocardiography can detect cardiac amyloidosis approximately 218 days before clinical diagnosis, offering a promising tool for earlier patient identification and timely treatment.

The Gist

Imagine your heart is a house, and over time, a sticky, invisible glue (called amyloid) starts to seep into the walls. This glue makes the walls thick and stiff, preventing the heart from pumping blood properly. This condition is called Cardiac Amyloidosis.

The problem is that this "glue" is a master of disguise. It looks a lot like other common heart problems, like high blood pressure or valve issues. Because of this, doctors often miss it for a long time, treating the wrong problem while the glue keeps building up. By the time they realize what's actually happening, the damage is often severe.

However, there is good news. New medicines exist that can stop the glue from building up, but they only work best if you catch the problem early.

The "Super-Scanner" AI

This paper is about a new tool: Artificial Intelligence (AI). Think of this AI as a super-powered security camera that has been trained to look at heart ultrasound videos (echocardiograms) and spot the tiny, early signs of this "glue" that human eyes might miss.

Here is what the researchers found:

1. The "Time Travel" Discovery
The researchers took a look back at heart scans of 349 patients who were eventually diagnosed with this condition. They ran these old scans through their AI "super-scanner."

• The Result: The AI flagged the problem about 7 to 9 months earlier than the human doctors did.
• The Analogy: Imagine you are driving a car, and the engine is making a weird noise. You don't realize it's serious until the car breaks down on the highway. This AI is like a mechanic who listens to the engine while you are still driving smoothly and says, "Hey, I hear a tiny rattle that means you're going to break down in six months. Let's fix it now."

2. The "Wrong Reason" Scans
The study found that most of these patients had already had heart scans done before they were diagnosed. But they didn't get the scans because they suspected amyloidosis. They got them because they had heart failure, chest pain, or irregular heartbeats.

• The Analogy: It's like someone going to the doctor because they have a cough, and the doctor finds a hidden tumor in their lung that they didn't know was there. The AI looked at these "cough" scans and realized, "Wait a minute, this heart looks like it has amyloidosis, even though the doctor was looking for something else."

3. Why This Matters
Because the AI can spot the disease months earlier, patients can get the new, life-saving medicines sooner.

• The Analogy: If you catch a fire when it's just a tiny spark, you can put it out with a cup of water. If you wait until it's a roaring inferno, you need a fire truck and a lot of damage is already done. This AI helps us find the "spark" before it becomes a "fire."

The Bottom Line

This research shows that we have a powerful new tool that can look at old heart scans and say, "This patient needs help now, not later." By using this AI, doctors might be able to catch this tricky disease much earlier, treat it faster, and help patients live longer, healthier lives.

Note: This study is a "preprint," meaning it is a new discovery that has been shared with the public but hasn't been fully checked by other experts yet. It's like a draft of a recipe that looks delicious but needs a final taste test before it's served to everyone.

Technical Summary

Based on the provided preprint, here is a detailed technical summary of the study "Artificial Intelligence Detects Cardiac Amyloidosis Before Clinical Diagnosis."

1. Problem Statement

• Clinical Challenge: Cardiac Amyloidosis (CA) is a progressive disease caused by amyloid fibril deposition in the myocardium. Despite the recent approval of novel disease-modifying treatments that can stabilize progression, diagnosis is frequently delayed.
• Root Causes of Delay:
  • Phenotypic Mimickers: CA often presents with symptoms and imaging features similar to other conditions (e.g., hypertensive heart disease, hypertrophic cardiomyopathy), leading to misdiagnosis.
  • Missed Opportunities: Patients frequently undergo echocardiography prior to a formal CA diagnosis, but these studies are often ordered for other indications (e.g., heart failure, valvular disease) rather than specifically for amyloidosis screening. Consequently, early signs of CA are overlooked by human readers.
• Objective: To determine if an Artificial Intelligence (AI) model applied to historical echocardiograms can detect CA significantly earlier than the current clinical standard of care.

2. Methodology

• Study Design: Retrospective cohort study conducted at Cedars-Sinai Medical Center (CSMC).
• Data Source:
  • Population: 396 patients with a clinical diagnosis of amyloidosis (with or without cardiac involvement) identified via ICD-9/10 codes and manual chart review between 2006 and 2024.
  • Sub-cohorts:
    • All-comers: 304 patients who had echocardiography studies performed prior to their initial clinical diagnosis.
    • Cardiac Amyloidosis (CA) Specific: 221 patients within the all-comer group confirmed to have cardiac involvement.
• AI Model:
  • An externally validated, open-source deep learning model (previously developed at Stanford Healthcare) was deployed.
  • Input: Echocardiography images (specifically Apical 4-chamber views implied by abbreviations).
  • Output: A probability score predicting the presence of cardiac amyloidosis.
  • Validation: The model was tested on the CSMC cohort, which had no overlap with the original development cohort.
• Analysis:
  • The AI model re-reviewed all echocardiograms performed prior to the clinical diagnosis date.
  • The study calculated the time interval between the first AI-positive scan and the actual clinical diagnosis.
  • Demographic and clinical characteristics (e.g., LVEF, wall thickness) were compared between patients with and without cardiac involvement.

3. Key Contributions

• Retrospective Validation: This study provides external validation of an AI model for CA detection in a real-world clinical setting distinct from its training environment.
• Temporal Analysis: It quantifies the "diagnostic gap," demonstrating that AI can identify CA months to years before a human clinician makes the diagnosis.
• Indication Agnosticism: The study highlights that AI can detect CA even when the echocardiogram was ordered for non-specific indications (e.g., heart failure, atrial fibrillation) rather than suspected amyloidosis.
• Differentiation of Subtypes: The analysis breaks down AI performance across different amyloidosis types (AL vs. ATTR vs. others).

4. Key Results

• Prevalence of Prior Imaging: 76.8% (304/396) of patients with amyloidosis had at least one echocardiogram performed before their clinical diagnosis.
• Diagnostic Delay:
  • The median time between the first echocardiogram and clinical diagnosis was 131 days.
  • AI Detection Lead Time:
    • In patients with confirmed Cardiac Amyloidosis (CA): AI was positive on average 218 days (approx. 7 months) prior to clinical diagnosis.
    • In All-comers (including non-cardiac amyloidosis): AI was positive on average 209 days prior.
    • Long-term Detection: 31.4% of CA patients were identified by AI at least 1 year before clinical diagnosis.
• Model Performance (Sensitivity):
  • CA Cohort: AI detected 90/221 (40.7%) of CA patients prior to diagnosis.
  • All-comer Cohort: AI detected 100/304 (32.9%) of all amyloidosis patients.
  • Subtype Breakdown (CA Cohort): AI was positive in 58.9% of AL amyloidosis cases and 37.8% of ATTR cases.
• Clinical Correlations:
  • Patients with CA had significantly increased interventricular septal (IVSd) and left ventricular posterior wall (LVPWd) thickness compared to those without cardiac involvement ($p < 0.001$).
  • LVEF was slightly lower in the CA group ($54.1\%$ vs $57.3\%$, $p=0.01$).
  • Imaging Indications: Only 7.7% of prior scans listed "possible amyloidosis" as the indication. The most common indications were heart failure (20.2%) and valvular disease (17.7%).
• Inflection Point: Model sensitivity showed an inflection point 3–6 months prior to diagnosis, where the proportion of positive cases stabilized to levels similar to those within one month of diagnosis.

5. Significance and Implications

• Early Intervention: Since disease-modifying treatments for CA are most effective when initiated early, reducing the diagnostic delay by ~7 months (via AI screening) could significantly improve patient outcomes and survival.
• Workflow Integration: The study suggests that integrating AI into routine echocardiography workflows could act as a "safety net," flagging high-risk patients who are being evaluated for other cardiac conditions.
• Cost-Effectiveness: By identifying high-probability cases earlier, healthcare systems can prioritize these patients for definitive (and often expensive) diagnostic testing (e.g., nuclear scintigraphy, biopsy, genetic testing) sooner.
• Future Directions: The authors propose that combining AI detection with clinical history and risk factors could further refine the specificity and utility of early CA screening.

Note: As this is a preprint (not yet peer-reviewed), the findings represent preliminary research and should not currently guide clinical practice without further validation.