As demand for physiological signal analysis grows in healthcare and wearable devices, building unified, standardized feature representations for continuous physiological signals such as ECG has become an important part of technical R&D. Ant inclusionAI recently released and open-sourced the physiological signal encoder LingSante-Pulse-1.0 on Hugging Face under the permissive MIT open-source license, providing developers with a foundational signal encoding asset.
Core positioning and prerequisites: provides pure signal representations and must be used with an inference code package
Functionally, LingSante-Pulse-1.0 is a physiological signal encoder whose primary role is to output signal representations. The model itself does not include prediction heads for specific downstream tasks.
Before using this model asset, developers need to install the standalone inference code package lingsante_pulse. After installation, calling LingSantePulse.from_pretrained() will automatically download the model on first use; to directly load assets already downloaded in the current model directory, use LingSantePulse.from_pretrained(".", local_files_only=True). In addition, users can lock a specific commit version via revision, or use local_files_only=True to load local and cached assets.
Input/output interface and preprocessing mechanism: supports multi-dimensional input with fixed 128-Hz processing
LingSante-Pulse-1.0 offers good flexibility in input formats, supporting [T], [C, T], [B, C, T], or an aligned channel dictionary (for example, passing the actual sample rate sample_rate and a channel description containing information such as "modality": "ECG", "lead": "II").
When using it, developers need to specify the actual sample rate and physical channel identity, while the accompanying inference code package automatically handles several low-level processing tasks:
Unified preprocessing and masking: the code package handles fixed 128-Hz preprocessing and also handles observed samples and storage-padding masks.
Feature output structure: the model returns CLS features and observed-valid PatchMean separately, both with dimension [B, C*256], and channels are concatenated in input order. When return_tokens=True is set, it can additionally return sequentially ordered patch features and a validity mask divided at 0.25-second intervals.
Repository files and channel vector provenance
The open-source repository contains complete model and channel configuration documentation files:
model.safetensors: stores encoder parameters.
config.json: records the model format and supports automatic file integrity checks.
channels.npz: stores fixed channel conditioning vectors and names.
channels.json: provides physical description information for channel identities.
Regarding the source of the fixed channel vectors, the official documentation states that these vectors are generated by BAAI/bge-small-en-v1.5 through CLS pooling and L2 normalization. The repository contains only the final generated vector results and does not distribute the upstream encoder code, weights, or tokenizer.


