You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

ATLAS-EHR

ATLAS-EHR is a PyTorch representation model for longitudinal structured EHR events. It returns per-timestamp z_set and z_context representations; it is research software, not a clinical device.

Release version: 0.2.0. The public runtime uses only JSON and safetensors—never Lightning checkpoints or research code.

Installation

Clone the intended Hugging Face repository and install its packaged runtime:

git clone https://huggingface.co/sunx999/ATLAS-EHR
cd ATLAS-EHR
python -m pip install .

This installs the declared PyTorch, NumPy, pandas, PyArrow, safetensors, and Hugging Face Hub dependencies. ATLAS-EHR is not published on PyPI; Lightning and the research src/ package are not required. After installation, load the model from the local checkout or with the Hub API below.

Input requirement

Users must supply already-prepared model-space events: ATLAS variable ID (VID), prepared value, and admission-relative time in minutes. The released schema is required. ATLAS-EHR does not extract raw EHR data, harmonize concepts, convert units, normalize values, or map local schemas.

Representation inference

from atlas_ehr import ATLASModel, AtlasCollator, PreparedSample

model = ATLASModel.from_pretrained_local("pretrained_backbone").eval()
sample = PreparedSample("example", [1, 2], [0.2, 1.0], [0.0, 30.0])
batch = AtlasCollator.from_pretrained_config("pretrained_backbone")([sample])
output = model(**batch)

For prepared Parquet data, use AtlasParquetDataset with a normal PyTorch DataLoader and AtlasCollator. model.encode(**batch) is an inference convenience method; model(**batch) remains differentiable. Freeze or unfreeze the backbone and attach a user-defined head for custom downstream work.

Formal task variants

Task artifacts are explicit variants under task_artifacts/variants/<task>/<source_name>. The checked-in release registry supplies one recommended release variant per task; the other source remains an available alternative. source_name (z_set or z_context) is artifact-defined and never selected from runtime metrics.

Task	Recommended variant
mortality	z_set
LOS ≥3d	z_set
LOS ≥7d	z_context
readmission	z_set
ICD	z_context
HCUP	z_context
DRG	z_set

For LOS ≥7d, z_context is the recommended MIMIC/formal-release variant; z_set remains available as an alternative.

from atlas_ehr import ATLASForPrediction

predictor = ATLASForPrediction.from_pretrained_local(
    "pretrained_backbone", task="mortality"
)
result = predictor.predict_samples([sample])

To request an available alternative explicitly, use task="mortality", source_name="z_context"; direct task_variant="task_artifacts/variants/mortality/z_context" loading remains supported.

Visibility is fixed by task: mortality first 48 h; los_ge_3d and los_ge_7d first 24 h; readmission, ICD, HCUP, and DRG full visible admission. Cutoffs are applied before contextual representation computation. ICD/HCUP use exported validation-fitted thresholds; DRG uses its competitive cardinality decoder.

Hugging Face Hub

model = ATLASModel.from_pretrained("owner/ATLAS-EHR", revision="v0.2.0")
predictor = ATLASForPrediction.from_pretrained(
    "owner/ATLAS-EHR", task="mortality", revision="v0.2.0"
)

Hub loading downloads declared files from one pinned revision and applies the same local schema, safetensors, and provenance validation.

Limitations

The pretrained model requires released-schema semantics and should not be assumed to transfer unchanged to arbitrary hospitals or raw EHR systems. Formal task variants are frozen-probe artifacts, not jointly fine-tuned clinical deployment models.

Downloads last month: -; Downloads are not tracked for this model. How to track