An ECG-based deep learning model detected ATTR-CM with AUROC between 0.916 and 0.990 across five hospital test sets, showing strong multi-hospital generalizability.
Can an ECG-based deep learning model accurately detect transthyretin amyloid cardiomyopathy (ATTR-CM)?
An ECG-based deep learning model demonstrated high accuracy (AUROC >0.91) for detecting ATTR-CM, highlighting its potential as an accessible screening tool for earlier diagnosis.
Absolute Event Rate: 0% vs 0%
Abstract Background Transthyretin amyloid cardiomyopathy (ATTR-CM) is an underdiagnosed, progressive disorder characterized by pathological deposition of misfolded transthyretin protein in the myocardium. Despite therapeutic options such as tafamidis, early detection remains difficult due to nonspecific clinical presentations. Recent machine learning models using electrocardiogram (ECG) have shown promising performance in detecting cardiac amyloidosis, highlighting the feasibility of AI-assisted detection for related cardiac conditions. Purpose This study aims to develop and validate a deep learning model for detecting ATTR-CM from 12-lead ECGs using multi-institutional data from an Asian population. Methods 12-lead ECGs were retrospectively collected from five university hospitals, including patients with ATTR-CM confirmed by nuclear scintigraphy (99mTc-DPD or 99mTc-PYP) or endomyocardial biopsy. For each positive patient, we selected ECGs recorded within six months prior to diagnosis. Negative controls were drawn from one hospital (Hospital A), with a whole-body bone scan showing no cardiac uptake and no ICD-10 codes for amyloidosis or infiltrative disease, with ECGs constrained to a 3-month window around the scan date. Vision Transformer models were pretrained via masked autoencoding and then fine-tuned on labeled ECG data. Model performance was assessed by calculating the area under the receiver operating characteristic curve (AUROC) on predefined hold-out test sets. Results A total of 59 patients with ATTR-CM (779 ECGs) and 46,275 controls (178,274 ECGs) were included. Five distinct hold-out test sets were constructed, each containing positive cases from a specific hospital and matched at a 1:10 ratio to negative controls based on age and sex. Analysis was performed using index ECGs (the first ECG within the six-month time window for positive cases, or the ECG closest to the WBBS date for negative cases). Across the five test sets, the model’s AUROC ranged from 0.916 (95% CI: 0.829–1.000) to 0.990 (95% CI: 0.970–1.000). This performance indicates robust generalizability despite negative samples being drawn from a single institution. Conclusions An ECG-based deep learning model demonstrated strong performance in detecting ATTR-CM across multiple hospitals, suggesting potential utility for earlier and more accessible identification of this underdiagnosed condition.
Oh et al. (Sat,) reported a other. An ECG-based deep learning model detected ATTR-CM with AUROC between 0.916 and 0.990 across five hospital test sets, showing strong multi-hospital generalizability.