Conference Coverage

Multimodal AI Flags ATTR-CM Risk With 85.6% Sensitivity Using Several Data Types

Edited by:
Brian McCurdy

Key Highlights

  • The system used data ending 6 months before each patient’s index date to assess potential need for technetium pyrophosphate scanning.
  • Validation compared ATTR-CM cases with age- and sex-matched patients with heart failure with preserved ejection fraction.
  • Overall accuracy was 80.2%, and precision was 77.3% in the balanced cohort.
  • Each screening flag included a report with reasoning traceable to patient-specific findings.

A multimodal artificial intelligence system identified transthyretin amyloid cardiomyopathy (ATTR-CM) risk with 85.6% sensitivity in a retrospective validation involving 500 Penn Medicine patients. The researchers presented the findings at the 2026 American Heart Association meeting on November 9.

Researchers at the University of Pennsylvania in Philadelphia developed the system to flag patients who might benefit from confirmatory technetium pyrophosphate (PYP) scanning. Their approach sought to integrate findings distributed across echocardiography scans (ECGs) and longitudinal clinical documentation, using only data available before diagnosis.

Deep learning models extracted quantitative features from ECGs, while reasoning large language models extracted findings from free-text reports and notes. The system organized those findings into a patient-level knowledge graph. An agentic reasoning layer then performed a general cardiac differential diagnosis followed by an ATTR-CM–specific assessment.

The retrospective, balanced cohort included patients identified using amyloidosis diagnostic codes and an active order for tafamidis, acoramidis, or vutrisiran. Controls were age- and sex-matched patients with heart failure with preserved ejection fraction (HFpEF) who had seen a general cardiologist and had no amyloidosis clinic encounter. Eligible data extended from 2015 at the earliest through 6 months before each patient’s index date.

Study Findings

The system achieved sensitivity of 85.6% and specificity of 74.8%. Precision was 77.3% and overall accuracy was 80.2%. Reported F1 score was 0.812 (95% CI, 0.771–0.850) and F3 score was 0.846 (95% CI, 0.806–0.882).

The investigators emphasized that specificity reflected discrimination against HFpEF controls, which they characterized as a particularly challenging clinical mimic, rather than healthy comparators. Each flagged assessment was accompanied by a report with traceable reasoning grounded in patient-specific portions of the knowledge graph. These results represent performance in the study’s retrospective validation cohort.

Clinical Implications

According to the investigators, the findings suggest that the system can identify relevant relationships across different clinical data types when assessing ATTR-CM risk. They stated that the knowledge graph allows auditing of findings, supporting clinical trust and workflow integration. They also described the platform as adaptable to additional modalities and other infiltrative cardiomyopathies.

Expert Commentary

“This multimodal agentic system demonstrates robust discrimination of ATTR-CM risk, suggesting capacity to identify relevant cross-modal relationships,” the investigators concluded.

Reference

Singhal P, Kim J, Kripke C, et al. A multimodal agentic knowledge graph system for screening undiagnosed ATTR-CM. Abstract MDP2528. Presented at: American Heart Association 2026; November 9, 2026.