Abstract Body (Do not enter title and authors here): Hypothesis and Purpose: Transthyretin amyloid cardiomyopathy (ATTR-CM) remains underdiagnosed despite therapies that can stabilize disease. We hypothesized that a machine learning model trained on ECG waveforms and TTE measurements could better identify patients at risk for undiagnosed ATTR-CM compared to existing risk scores, and that integrating this model into a real-world, AI-augmented clinical workflow would increase the rate of ATTR-CM detection and yield of testing.

Study Design and Methods: We developed a machine learning model, ATTRACTnet, using electrocardiogram (ECG) waveforms and transthoracic echocardiogram (TTE) measurements, and conducted a multicenter, single-arm, open-label trial to evaluate its real-world performance (NCT06469372). The trial spanned eight hospitals in New York City, where eligible patients were identified by the model, notified, and offered diagnostic testing and follow-up care.

Sample Size: 799 patients contributed 22,344 ECG-TTE pairs to model development. The prospective trial enrolled 256 eligible patients, of whom 50 underwent cardiac amyloidosis testing.

Population Studied: Patients ≥ 50 years old with an ECG and TTE within 2 years with left ventricular (LV) wall thickness ≥ 12 mm who had an AI model score of ≥0.5 were eligible for inclusion in the trial. Exclusion criteria included prior ATTR-CM testing, heart transplantation, dementia, expected lifespan <1 year, hypertrophic cardiomyopathy, or LV wall thickness <14 mm explained by uncontrolled hypertension or moderate/severe aortic stenosis.

Intervention(s): Clinicians of flagged patients were contacted via EHR messaging. Upon approval, patients were offered cardiac amyloidosis scintigraphy and monoclonal protein testing at Columbia or local sites per clinician discretion.

Power Calculations: This feasibility trial used prespecified interim analyses after 25 and 50 patients.

Primary End Points: Proportion of tested patients diagnosed with ATTR-CM vs. historical and contemporary controls.

Outcome(s)[Statistical Plan or Main Results]: ATTRACTnet outperformed a validated risk score (AUROC 0.85 vs. 0.59). Of 50 tested, 24 (48%) were diagnosed with ATTR-CM, exceeding historical (15.3%) and contemporary (17.0%) rates (p<0.001), and 88% of diagnosed patients began therapy within 3 months. The model flagged patients a median of 345 days before enrollment. The trial resulted in an 18% relative increase in system-wide diagnoses compared to the prior year.