Transforming ECG Images to Waveforms with UNet-LSTM and Multilabel Classification with ResNet-LSTM: A Multimodal Approach by PulsePlex

In the 2024 George B. Moody PhysioNet/Computing in Cardiology Challenge, our team, PulsePlex, worked on digitizing and classifying electrocardiograms (ECGs) from images and paper printouts, aiming to enhance cardiac care accessibility in underserved regions.For the digitization task, we developed a hybrid model combining a UNet and LSTM to convert ECG images into waveform representations.Our model achieved a leaderboard score of 0.081 on hidden validation data and a final test signal-tonoise ratio (SNR) of -0.247 0.096, securing 11th place out of 16 teams.For classification, we used a multimodal fusion approach, with a ResNet extracting image features and an LSTM processing the digitized ECG signals for multilabel classification of cardiac conditions.This resulted in a leaderboard score of 0.317 on hidden validation data and a test macro F-measure of 0.222 0.000 on hidden test data, placing us 10th out of 16 teams in the Official Rankings.These outcomes underscore the challenges of ECG digitization and classification from nondigital sources while highlighting the promise of multimodal approaches in advancing automated ECG analysis.