Improving transthyretin cardiac amyloidosis detection from electrocardiograms through the Willem artificial intelligence platform.

BACKGROUND

Transthyretin amyloid cardiomyopathy (ATTR-CM) is a frequently underdiagnosed disease in which delay in diagnosis limits the efficacy of treatments. Artificial intelligence (AI) applied to standard 12-lead electrocardiograms (ECGs) is promising and may help improve ATTR-CM early detection.

OBJECTIVES

The goals of this study were to extend the Willem AI cloud platform for ATTR-CM detection from 12-lead ECGs and to evaluate its performance across relevant patient subgroups, including early-stage disease.

METHODS

This was a single-center retrospective study of adults evaluated for ATTR-CM. Controls had negative 3,3-diphosphono-1,2-propanodicarboxylic acid scintigraphy and clinical assessment to exclude ATTR-CM. After quality control, 20,754 ECGs from 2901 individuals [585 patients with ATTR-CM (20.2%) and 2316 controls (79.8%)] were used. Patients were split into training/validation and test data sets, with stratification by amyloidosis status, sex, and age groups. A time-slicing convolutional neural network generated ECG probabilities aggregated by max-pooling. Threshold selection targeted 90% sensitivity in validation.

RESULTS

In the test cohort, Willem AI achieved an area under the receiver operating characteristic curve (AUC) of 0.88 (95% confidence interval, 0.85-0.91), a sensitivity of 80.7%, and a specificity of 78.5%. Performance was similar for hereditary transthyretin amyloidosis or hereditary ATTR (ATTRv; AUC, 0.91) and wild-type transthyretin amyloidosis or wild type ATTR (ATTRwt; AUC, 0.88) and remained informative in early presentations (asymptomatic sensitivity 68.4%; New York Heart Association class I 73.9%). Willem AI achieved high accuracy in "red-flag" scenarios: carpal tunnel syndrome (AUC, 0.92), lumbar spinal stenosis (AUC, 0.92), and polyneuropathy (AUC, 0.91).

CONCLUSION

This work expands Willem AI capabilities, enabling automated ATTR-CM detection from routine 12-lead ECGs in a real-world referral population. The Willem AI model may support informed clinical management and scalable triage to trigger confirmatory testing and accelerate diagnosis.