Poster #48: Pankhuri Singhal PhD

A Multimodal Agentic Knowledge Graph System for Screening Undiagnosed Transthyretin Amyloid Cardiomyopathy

Pankhuri Singhal, PhD1; Jenny Kim, PhD1; Colleen Kripke, DPT1; Marjarie Risman, MA1,2; Michael Levin, MD3; Emeka C Anyanwu, MD3; Walter R. Witschey, PhD4; Daniel J. Rader, PhD1,2; Pranav Chandrashekhar, MD3; Anurag Verma, PhD1,2 {1Division of Translational Medicine and Human Genetics, Department of Medicine, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA; 2Department of Genetics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA; 3Division of Cardiology, Department of Medicine, University of Pennsylvania, Philadelphia, PA; 4Division of Radiology, University of Pennsylvania, Philadelphia, PA)

Transthyretin amyloid cardiomyopathy (ATTR-CM) is a progressive, underdiagnosed infiltrative cardiomyopathy with disease-modifying therapies. Diagnosis is often delayed because of heterogeneous findings, including increased ventricular wall thickness, low ECG voltage, pseudoinfarct patterns, carpal tunnel syndrome, and spinal stenosis. We developed and validated a multimodal agentic AI system to identify patients who may benefit from technetium pyrophosphate (PYP) scanning for undiagnosed ATTR-CM. Hypothesis: An agentic AI system integrating echocardiograms, echocardiography reports, electrocardiogram (ECG) waveforms and reports, and longitudinal clinical notes could accurately identify patients warranting PYP scanning using only prediagnosis data. Methods: The system used deep learning to extract quantitative echocardiographic and ECG features and tool-calling large language models to extract clinical findings from notes. It constructed patient-level knowledge graphs encoding cross-modal, temporal, and associative relationships. An agentic reasoning layer performed disease-agnostic cardiac differential diagnosis with anti-confirmation-bias guardrails. An LLM-as-judge framework validated intermediate inferences and conclusions against empirical patient graphs. A clinician interface enabled interactive graph and report review. We retrospectively evaluated a balanced cohort of 500 Penn Medicine patients. Cases with amyloidosis codes (E85.0, E85.2–E85.9) were confirmed by active tafamidis, acoramidis, or vutrisiran orders. Controls were age- and sex-matched patients with HFpEF, general cardiology care, and no amyloidosis clinic encounter. Data were limited to 2015 through six months before each patient’s index date. Results: Sensitivity was 0.856 (95% CI, 0.807–0.899), specificity 0.748 (0.690–0.802), precision 0.773 (0.718–0.826), F1 0.812 (0.771–0.850), F3 0.846 (0.806–0.882), and accuracy 0.802 (0.762–0.840). Specificity reflected discrimination from HFpEF, a challenging clinical mimic. Each flag included a structured report with traceable, patient-specific reasoning. Conclusion: This system identified ATTR-CM risk using exclusively prediagnosis data, potentially enabling earlier PYP evaluation. Its auditable knowledge graph supports clinical trust and workflow integration. The disease-agnostic, modality-extensible architecture may apply to other underdiagnosed cardiomyopathies.