Longer Temporal Context Improves ECG AI Model Performance.

Research into self-supervised electrocardiogram (ECG) models often uses short signal segments and tokenized data, but the impact of these choices on clinical utility has been unclear. A new controlled study explored how varying the input horizon (from 16 seconds to 10 minutes) and the data representation (continuous convolutional patch embeddings versus fixed vector-quantized tokens) affects model performance. The findings indicate that providing more temporal context, specifically 5 to 10 minutes of ECG data, leads to stronger transfer learning capabilities and higher accuracy in patient-level retrieval tasks. This suggests that longer contexts are better at capturing subtle, slow-varying rhythm dynamics and individual patient characteristics. Furthermore, the study found that continuous patch embeddings consistently outperformed discretized tokens across all tested horizons. This implies that the process of quantizing ECG signals into tokens might discard crucial waveform details that are clinically relevant, advocating for continuous encoders in future ECG foundation models.