Development, validation, and evaluation of AI model to detect ATTR-CM

ATTRACTnet was developed as an AI-based screening tool for transthyretin-mediated amyloid cardiomyopathy (ATTR-CM). In a real-world population, use of the ATTRACTnet score resulted in a higher diagnostic yield compared with regular screening.

This summary is based on the publication of Jain SS, Sun T, Pierson E, et al. - Detecting Transthyretin Cardiac Amyloidosis With Artificial Intelligence: A Nonrandomized Clinical Trial. JAMA Cardiol. 2026 Feb 1;11(2):117-124. doi: 10.1001/jamacardio.2025.4591

Introduction and methods

Background

Transthyretin-mediated amyloid cardiomyopathy (ATTR-CM) is a progressive and fatal disease characterized by the deposition of amyloid fibrils, which consist of misfolded transthyretin (TTR) aggregates, in the myocardium. As the initial symptoms and signs of ATTR-CM are nonspecific, early-stage ATTR-CM has historically been underdiagnosed [1,2], particularly in Black patients [3,4]. Given the rapid expansion of treatment options for ATTR-CM, early diagnosis is imperative to delay disease progression and improve patient outcomes [5-7]. However, traditional risk scores and AI models have not been used to prospectively assess the impact of AI-augmented diagnostic processes on improving the detection of ATTR-CM in a real-world setting [8-10].

Aims of the study

The study aims were to (1) develop and externally validate an AI model (ATTRACTnet) and assess whether ATTRACTnet has additional benefit over a validated risk stratification method (i.e., ATTR-CM risk score) and (2) prospectively evaluate an AI-augmented workflow for ATTR-CM screening to assess the impact of identifying undiagnosed ATTR-CM in a real-world population.

Methods

The researchers constructed ATTRACTnet using data on ECG waveforms, echocardiographic measurements, demographics, and diagnosis codes for orthopedic manifestations of amyloidosis from 799 patients who were seen at the New York–Presbyterian Hospital and Columbia University Irving Medical Center in the period 2010–2023. Performance of ATTRACTnet was assessed using 5-fold cross validation for, among others, the area under the receiver operating characteristic curve (AUROC). External validation involved an outside dataset of patients who underwent amyloid scintigraphy at the New York–Presbyterian Hospital/Weill Cornell Medical Center in the period 2014–2023 (n=422).

To evaluate the real-world performance of the AI-augmented workflow, the Cardiac Amyloidosis Discovery Trial was conducted, a single-system, multicenter, open-label, single-arm, nonrandomized clinical trial. Patients aged ≥50 years with LV wall thickness ≥12 mm (or ≥14 mm in case of uncontrolled hypertension or moderate or severe aortic stenosis) and ATTRACTnet score ≥0.5 were potentially eligible (n=1471). Exclusion criteria included suspected ATTR-CM in clinical notes, prior ATTR-CM testing, diagnosis of hypertrophic cardiomyopathy, nursing facility residence, advanced dementia, and expected life span <1 year. Eligible patients were offered nuclear scintigraphy testing, monoclonal protein testing, and follow-up care on agreement from the treating physician.

The percentage of patients diagnosed with ATTR-CM in the study population was compared with that of historical controls evaluated in the previous 12 months (n=887) and contemporary controls evaluated during the same period as the study population (n=854), at the same clinical centers. The current analysis comprised the interim results of this trial (May 2024–May 2025).

Outcome

The primary endpoint was a diagnosis of ATTR-CM based on consensus criteria, including the presence or absence of myocardial uptake on single-photon emission CT imaging.

Main results

Validation and testing of ATTRACTnet

• In 5-fold cross-validation analysis, ATTRACTnet had a sensitivity of 0.56, specificity of 0.92, negative predictive value of 0.84, and positive predictive value of 0.74.
• ATTRACTnet showed good discrimination for ATTR-CM detection in the internal test set (AUROC: 0.85; 95%CI: 0.77–0.85) and external test set (AUROC: 0.82; 95%CI: 0.81–0.83), with no differences between racial and ethnic subgroups.
• To investigate the number of patients across the system with possible elevated ATTR-CM risk and how this number can be decreased by AI assessment, the ATTR-CM risk score was applied to patients aged ≥50 years with LV wall thickness ≥12 mm who did not have amyloid scintigraphy in the period 2010–2023 (n=14,585). In this subgroup, 9892 patients (67.8%) were identified with the ATTR-CM risk score as having elevated ATTR-CM risk and only 1515 patients were tested for ATTR-CM in routine clinical care. Of these 9892 patients, 1196 (12.1%) had an ATTRACTnet score ≥0.5.

Cardiac Amyloidosis Discovery Trial

• At the time of the interim analysis, 50 patients had undergone cardiac amyloid testing after physician and patient approval, 24 (48.0%; 95%CI: 34.8%–61.5%) of whom tested positive for ATTR-CM. Within 3 months of diagnosis, 21 patients (88%) were prescribed ATTR-CM treatment.
• The rate of ATTR-CM diagnoses was lower among historical controls (15.3%; 95%CI: 13.1%–17.9%; P<0.001).
• The positivity rate was also lower in the contemporary control group (17.0%; 95%CI: 14.6%–19.6%; P<0.001). There was no significant difference between the rate of positive tests in the historical and contemporary control groups.

Conclusion

The newly developed AI model ATTRACTnet was validated in internal and external test sets. In a pilot study, the Cardiac Amyloidosis Discovery Trial, screening for ATTR-CM based on the ATTRACTnet score resulted in a higher diagnostic yield compared with regular screening in historical and contemporary control groups. According to the authors, “AI-augmented screening may improve ATTR-CM detection and identify patients who are missed by usual care.”

Find this article online at JAMA Cardiol.

References

1. Rapezzi C, Lorenzini M, Longhi S, et al. Cardiac amyloidosis: the great pretender. Heart Fail Rev. 2015;20(2):117-124. doi: 10.1007/s10741-015-9480-0
2. Rozenbaum MH, Large S, Bhambri R, et al. Impact of delayed diagnosis and misdiagnosis for patients with transthyretin amyloid cardiomyopathy (ATTR-CM): a targeted literature review. Cardiol Ther. 2021;10(1):141-159. doi: 10.1007/s40119-021-00219-5
3. Shah KB, Mankad AK, Castano A, et al. Transthyretin cardiac amyloidosis in Black Americans. Circ Heart Fail. 2016;9(6):e002558. doi: 10.1161/CIRCHEARTFAILURE.115.002558
4. Spencer-Bonilla G, Njoroge JN, Pearson K, Witteles RM, Aras MA, Alexander KM. Racial and ethnic disparities in transthyretin cardiac amyloidosis. Curr Cardiovasc Risk Rep. 2021;15(6):8. doi: 10.1007/s12170-021-00670-y
5. Maurer MS, Schwartz JH, Gundapaneni B, et al; ATTR-ACT Study Investigators. Tafamidis treatment for patients with transthyretin amyloid cardiomyopathy. N Engl J Med. 2018;379(11):1007-1016. doi: 10.1056/NEJMoa1805689
6. Rettl R, Mann C, Duca F, et al. Tafamidis treatment delays structural and functional changes of the left ventricle in patients with transthyretin amyloid cardiomyopathy. Eur Heart J Cardiovasc Imaging. 2022;23(6):767-780. doi: 10.1093/ehjci/jeab226
7. Fontana M, Berk JL, Gillmore JD, et al; HELIOS-B Trial Investigators. Vutrisiran in patients with transthyretin amyloidosis with cardiomyopathy. N Engl J Med. 2024;392(1):33-44. doi: 10.1056/NEJMoa2409134
8. Davies DR, Redfield MM, Scott CG, et al. A simple score to identify increased risk of transthyretin amyloid cardiomyopathy in heart failure with preserved ejection fraction. JAMA Cardiol. 2022;7(10):1036-1044. doi: 10.1001/jamacardio.2022.1781
9. Huda A, Castaño A, Niyogi A, et al. A machine learning model for identifying patients at risk for wild-type transthyretin amyloid cardiomyopathy. Nat Commun. 2021;12(1):2725. doi: 10.1038/s41467-021-22876-9
10. Goto S, Mahara K, Beussink-Nelson L, et al. Artificial intelligence-enabled fully automated detection of cardiac amyloidosis using electrocardiograms and echocardiograms. Nat Commun. 2021;12(1):2726. doi: 10.1038/s41467-021-22877-8