Applying masked autoencoder-based self-supervised learning for high-capability vision transformers of electrocardiographies

• DOI: 10.1371/journal.pone.0307978
• Publication Date: 2024-08-14T17:21:58Z
• Authors: Shinnosuke Sawano; Satoshi Kodera; Naoto Setoguchi; Kengo Tanabe; Shunichi Kushida; Junji Kanda; Mike Saji; Mamoru Nanasato; Hisataka Maki; Hideo Fujita; Nahoko Kato; Hiroyuki Watanabe; Minami Suzuki; Masao Takahashi; Naoko Sawada; Masao Yamasaki; Masataka Sato; Susumu Katsushika; Hiroki Shinohara; Norifumi Takeda; Katsuhito Fujiu; Masao Daimon; Hiroshi Akazawa; Hiroyuki Morita; Issei Komuro
• Keywords: Autoencoder; Benchmark (surveying)

The generalization of deep neural network algorithms to a broader population is an important challenge in the medical field. We aimed apply self-supervised learning using masked autoencoders (MAEs) improve performance 12-lead electrocardiography (ECG) analysis model limited ECG data. pretrained Vision Transformer (ViT) models by reconstructing data with MAE. fine-tuned this MAE-based on ECG-echocardiography from University Tokyo Hospital (UTokyo) for detection left ventricular systolic dysfunction (LVSD), and then evaluated it multi-center external validation seven institutions, employing area under receiver operating characteristic curve (AUROC) assessment. included 38,245 pairs UTokyo 229,439 all institutions. performances were significantly higher than that other Deep Neural Network across cohorts (AUROC, 0.913–0.962 LVSD, p < 0.001). Moreover, we also found improvements depending capacity amount training Additionally, maintained high even benchmark dataset (PTB-XL). Our proposed method developed