The Willem AI platform achieved an AUROC of 0.841 for identifying patients with transthyretin cardiac amyloidosis from 12-lead ECGs in a large dataset.
Does a deep-learning model based on the Willem AI platform accurately identify patients with ATTR-CA from standard 12-lead ECGs?
A deep-learning model applied to standard 12-lead ECGs demonstrated good discriminative power (AUROC 0.841) for identifying transthyretin cardiac amyloidosis.
Absolute Event Rate: 0% vs 0%
Abstract Objectives Misdiagnosis and delay in diagnosis are frequent in patients with transthyretin cardiac amyloidosis (ATTR-CA). Methods from simple and accessible tests such as ECG are needed to better identify and treat patients. We sought to develop and optimize a deep-learning model based on the Willem AI platform to identify patients with ATTR-CA from 12-lead electrocardiograms (ECGs). Methods We analyzed ECGs from all subjects undergoing 3,3-diphosphono-1,2-propanodicarboxylic acid (DPD) scintigraphy for routine care assessment of ATTR-CA at Site01 from 2009 to 2023. 523 patients exhibited cardiac uptake and after complete assessment were diagnosed with ATTR-CA, whereas 2117 patients without DPD uptake constituted the control group. The Willem AI platform was optimized to discriminate between ATTR-CA patients and controls from 10 seconds 12-lead ECGs. A total of 9183 ECG records from all recruited subjects were used for model development and evaluation (ATTR-CA: 5237, 57.03%; controls: 3946, 42.97%), considering ECGs independently. 3707 ECGs from ATTR-CA and 2753 ECGs from controls were used for training and validation (70.30% of the total), and 1530 ECGs from ATTR-CA and 1193 ECGs from controls for testing (29.70%). The model generated its binary output based on features learned from ECG signals in the training dataset. Each ECG signal was normalized on a per-lead basis. Class weighting technique was used to adjust the loss function to cope with class imbalance, and the area under the Area Under the Receiver Operating Characteristic curve (AUROC) was considered as the main evaluation metric. Results In the test dataset, AUROC for correct ATTR-CA classification was 0.841 (CI 95%: 0.826–0.856). Other performance metrics are represented in Table 1. The ROC curve is displayed in Figure 1. These results suggest the model has a good discriminative power to identify patients with ATTR-CA. Conclusions Willem AI platform identifies ATTR-CA patients using standard 12-lead ECGs with positive preliminary performance in this single-center study with a large dataset. Future work includes external validation in an ongoing multicenter european study to further evaluate model robustness and generalizability.Table 1
Gonzalez-Lopez et al. (Thu,) reported a other. The Willem AI platform achieved an AUROC of 0.841 for identifying patients with transthyretin cardiac amyloidosis from 12-lead ECGs in a large dataset.