This paper introduces ST-MEM (Spatio-Temporal Masked Electrocardiogram Modeling), a self-supervised learning framework designed to capture the spatio-temporal relationships in electrocardiogram (ECG) signals. The primary goal is to learn general representations of ECG data, which can be used for various disease screening tasks, especially when labeled data are scarce. ST-MEM employs a masked autoencoder (MAE) architecture, which reconstructs masked 12-lead ECG data, capturing both spatial and temporal information. The model includes lead-wise shared decoders, learnable lead embeddings, and separation embeddings to handle different lead combinations effectively. Extensive experiments demonstrate that ST-MEM outperforms other SSL baseline methods in arrhythmia classification tasks, both in standard 12-lead and reduced lead settings. The code for ST-MEM is available at <https://github.com/bakqui/ST-MEM>.This paper introduces ST-MEM (Spatio-Temporal Masked Electrocardiogram Modeling), a self-supervised learning framework designed to capture the spatio-temporal relationships in electrocardiogram (ECG) signals. The primary goal is to learn general representations of ECG data, which can be used for various disease screening tasks, especially when labeled data are scarce. ST-MEM employs a masked autoencoder (MAE) architecture, which reconstructs masked 12-lead ECG data, capturing both spatial and temporal information. The model includes lead-wise shared decoders, learnable lead embeddings, and separation embeddings to handle different lead combinations effectively. Extensive experiments demonstrate that ST-MEM outperforms other SSL baseline methods in arrhythmia classification tasks, both in standard 12-lead and reduced lead settings. The code for ST-MEM is available at <https://github.com/bakqui/ST-MEM>.