AI-enabled ECG imaging for the detection of cardiac amyloidosis across amyloidosis subtypes
A Achten, A Sheikh, E K Oikonomou, P M Croon, Y Razvi, J Mansell, A Porcari, L Venneri, A Martinez-Naharro, P N Hawkins, J D Gillmore, R Khera, M FontanaAbstract
Introduction
Diagnosing cardiac amyloidosis remains challenging due to non-specific clinical features which overlap with causes of heart failure. Artificial intelligence (AI)-enabled electrocardiography (ECG) may provide a scalable strategy for detecting cardiac involvement in amyloidosis. We hypothesized that myocardial remodelling associated with cardiac amyloidosis is detectable on 12-lead ECG images using artificial intelligence, enabling accurate discrimination between amyloidosis patients with and without cardiac involvement.
Methods
This retrospective study included 3,902 patients referred for suspected amyloidosis to a tertiary referral centre. Individuals without amyloidosis or without evidence of cardiac involvement served as the control group, while patients with amyloidosis and cardiac involvement were considered cases. Cardiac involvement was confirmed based on (i) myocardial biopsy, (ii) 99mtechnetium labelled 3,3-diphosphono-1,2-propanodicarboxylic acid scintigraphy or (iii) elevated T1 mapping and extracellular volume on cardiovascular magnetic resonance. A previously developed image-based artificial intelligence electrocardiography (AI-ECG) model was applied to detect cardiac amyloidosis, and its discriminatory performance was evaluated across predefined patient subgroups according to amyloidosis subtype (light chain vs transthyretin). Diagnostic performance for the detection of cardiac involvement was assessed using receiver operating characteristic curve analysis.
Results
Ultimately, 1301 patients were diagnosed with light chain amyloidosis (AL), 1913 with transthyretin amyloidosis (ATTR), 51 with serum A protein amyloidosis (AA) and amyloidosis was not present in 637 patients. Overall 2595 patients had amyloidosis with cardiac involvement (932 AL, 1661 ATTR, 2 AA), 670 patients had amyloidosis without cardiac involvement of whom 161 were TTR variant carriers. Patients with cardiac involvement were significantly older and more likely male (Table 1). In this study, the AI-ECG for cardiac involvement had an area under the curve (AUC) of 0.89 (95% CI: 0.87-0.90) for the overall cohort. Model performance remained robust across amyloidosis subtypes, with consistent discrimination between cardiac and non-cardiac involvement. The AUC was the highest when predicting cardiac involvement in variant ATTR patients (AUC 0.93, 95% CI: 0.87-0.94) as shown in Figure 1.
Conclusion
AI-enabled analysis of 12-lead ECG images demonstrated high diagnostic accuracy for identifying cardiac involvement in amyloidosis in a large cohort, with particularly strong performance in variant ATTR carriers. This scalable, non-invasive approach has the potential to support earlier diagnosis, risk stratification, and longitudinal monitoring of individuals at risk for amyloid cardiomyopathy.Receiver operating characteristic curvesFor image description, please refer to the figure legend and surrounding text.For image description, please refer to the figure legend and surrounding text.