New AI Model Detects Cardiac Amyloidosis Earlier With 93% Sensitivity
A multimodal AI tool combining echo data, lab work, and clinical markers outperforms older models at catching a deadly, underdiagnosed heart condition.
Summary
Cardiac amyloidosis — a serious heart condition caused by misfolded protein deposits — is notoriously hard to catch early, with diagnosis often delayed by a year. Researchers have developed a new AI model called AI-ECM that combines echocardiogram data with routine lab results and patient demographics to detect the disease more accurately. In an internal validation study, AI-ECM achieved 93% sensitivity and an AUC of 0.94, significantly outperforming the FDA-cleared TTE-only model it was compared against. The tool eliminated indeterminate classifications entirely, which could reduce diagnostic limbo for patients. Earlier detection means earlier access to treatments that can meaningfully slow disease progression.
Detailed Summary
Cardiac amyloidosis (CA) is a progressive and often fatal cause of heart failure, driven by misfolded protein deposits accumulating in the heart muscle. Despite growing recognition of the disease, diagnosis is frequently delayed by about a year after clinical suspicion — partly because no reliable frontline screening test exists. Even echocardiography, a standard cardiac imaging tool, struggles with accuracy due to CA's overlap with other heart conditions. This diagnostic gap has real consequences: the earlier CA is caught, the more effective available therapies tend to be.
Researchers at MedStar Health Research Institute have developed AI-ECM, a multimodal AI algorithm that layers demographics, laboratory biomarkers, and transthoracic echocardiography (TTE) parameters into a single diagnostic model. The tool builds on Us2.Ca, an FDA-cleared deep-learning model based solely on TTE data. By adding clinical and lab inputs, the team aimed to capture diagnostic signals that imaging alone misses.
In an internal validation study using multiethnic registry data, AI-ECM demonstrated meaningful improvements: sensitivity rose from 76% to 93%, and the area under the curve improved from 0.89 to 0.94. Crucially, the model eliminated indeterminate classifications entirely — a persistent problem with existing tools. It also showed particular gains in detecting light chain CA, a historically harder-to-identify subtype.
The tradeoff is a modest drop in specificity, from 91% to 85%, meaning slightly more false positives. However, editorial commentators from Brigham and Women's Hospital noted that for a screening tool designed to reduce missed diagnoses, higher sensitivity is the more clinically meaningful metric.
The model relies entirely on routine data already collected during early CA evaluation, making it practically scalable without additional testing burden. Prospective validation across broader populations is still needed before clinical deployment, but the findings represent a significant step toward AI-guided early detection of a condition where timing is everything.
Key Findings
- AI-ECM achieved 93% sensitivity vs 76% for the TTE-only model, dramatically reducing missed diagnoses.
- AUC improved from 0.89 to 0.94, reflecting meaningfully better overall diagnostic accuracy.
- The new model eliminated indeterminate classifications entirely, reducing diagnostic uncertainty for patients.
- AI-ECM showed improved detection of light chain CA, a harder-to-diagnose subtype.
- The model uses only routine clinical data already collected during standard CA workups, enabling scalable deployment.
Methodology
This is a news report summarizing a peer-reviewed study published in Circulation: Cardiovascular Imaging, a credible cardiology journal. The evidence basis is an internal validation study using multiethnic registry data; no randomized trial or prospective external validation has yet been conducted. An accompanying editorial from Harvard-affiliated clinicians provides independent expert commentary.
Study Limitations
The study is an internal validation only; external prospective validation across independent populations has not yet been completed. The modest drop in specificity means some false positives are expected, which could lead to unnecessary follow-up testing. Full clinical deployment details, including regulatory clearance, are not yet established.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.