This paper introduces ST-MEM (Spatio-Temporal Masked Electrocardiogram Modeling), a self-supervised learning framework designed to learn spatio-temporal features from 12-lead ECG data by reconstructing masked ECG signals. The framework incorporates spatio-temporal relationships inherent in ECG signals, which are crucial for accurate diagnosis and classification of heart diseases. ST-MEM outperforms other SSL baseline methods in various experimental settings for arrhythmia classification tasks and is adaptable to various lead combinations. Through quantitative and qualitative analysis, the paper demonstrates the spatio-temporal relationship within ECG data. The code for ST-MEM is available at https://github.com/bakqui/ST-MEM. The paper discusses the challenges of using ECG data for disease screening due to limited labeled data. Self-supervised learning (SSL) is a well-known approach to overcome this challenge, but naive application of SSL to ECG data without considering spatio-temporal relationships may yield suboptimal results. The paper proposes ST-MEM, which uses a spatio-temporal masked auto-encoder (MAE) architecture to learn general representations of ECG data. The framework includes a lead-wise shared decoder, learnable lead embeddings, and separation embedding to capture spatio-temporal relationships in ECG signals. The paper evaluates the performance of ST-MEM on various datasets, including pre-training datasets such as Chapman, Ningbo, and CODE-15, and downstream datasets such as PTB-XL, CPSC2018, and PhysioNet2017. The results show that ST-MEM achieves superior performance in arrhythmia classification tasks compared to other SSL methods, including contrastive and generative learning methods. The framework is also shown to be robust in low-resource settings and reduced lead scenarios, demonstrating its effectiveness in capturing spatio-temporal relationships in ECG data. The paper also discusses the effectiveness of general ECG representation learning in low-resource settings and reduced lead scenarios. The results show that ST-MEM outperforms other methods in these scenarios, demonstrating the importance of spatio-temporal information in ECG representation learning. The paper also presents an ablation study to evaluate the effectiveness of the lead indication module in ST-MEM, showing that it significantly improves performance in arrhythmia classification tasks. Finally, the paper discusses the interpretation of ST-MEM by analyzing self-attention mechanisms, showing that the framework effectively captures both spatial and temporal relationships in ECG signals. The results demonstrate that ST-MEM is a robust and effective framework for learning general ECG representations, which can be applied to various ECG problems, including arrhythmia classification and myocardial infarction detection. The paper concludes that ST-MEM provides a general ECG representation that encapsulates spatio-temporal relationships, which can benefit the healthcare industry.This paper introduces ST-MEM (Spatio-Temporal Masked Electrocardiogram Modeling), a self-supervised learning framework designed to learn spatio-temporal features from 12-lead ECG data by reconstructing masked ECG signals. The framework incorporates spatio-temporal relationships inherent in ECG signals, which are crucial for accurate diagnosis and classification of heart diseases. ST-MEM outperforms other SSL baseline methods in various experimental settings for arrhythmia classification tasks and is adaptable to various lead combinations. Through quantitative and qualitative analysis, the paper demonstrates the spatio-temporal relationship within ECG data. The code for ST-MEM is available at https://github.com/bakqui/ST-MEM. The paper discusses the challenges of using ECG data for disease screening due to limited labeled data. Self-supervised learning (SSL) is a well-known approach to overcome this challenge, but naive application of SSL to ECG data without considering spatio-temporal relationships may yield suboptimal results. The paper proposes ST-MEM, which uses a spatio-temporal masked auto-encoder (MAE) architecture to learn general representations of ECG data. The framework includes a lead-wise shared decoder, learnable lead embeddings, and separation embedding to capture spatio-temporal relationships in ECG signals. The paper evaluates the performance of ST-MEM on various datasets, including pre-training datasets such as Chapman, Ningbo, and CODE-15, and downstream datasets such as PTB-XL, CPSC2018, and PhysioNet2017. The results show that ST-MEM achieves superior performance in arrhythmia classification tasks compared to other SSL methods, including contrastive and generative learning methods. The framework is also shown to be robust in low-resource settings and reduced lead scenarios, demonstrating its effectiveness in capturing spatio-temporal relationships in ECG data. The paper also discusses the effectiveness of general ECG representation learning in low-resource settings and reduced lead scenarios. The results show that ST-MEM outperforms other methods in these scenarios, demonstrating the importance of spatio-temporal information in ECG representation learning. The paper also presents an ablation study to evaluate the effectiveness of the lead indication module in ST-MEM, showing that it significantly improves performance in arrhythmia classification tasks. Finally, the paper discusses the interpretation of ST-MEM by analyzing self-attention mechanisms, showing that the framework effectively captures both spatial and temporal relationships in ECG signals. The results demonstrate that ST-MEM is a robust and effective framework for learning general ECG representations, which can be applied to various ECG problems, including arrhythmia classification and myocardial infarction detection. The paper concludes that ST-MEM provides a general ECG representation that encapsulates spatio-temporal relationships, which can benefit the healthcare industry.