ACLF browser report models
ONNX report-processing models for the browser-based ACLF prognosis research prototype. These files are downloaded by the website and executed locally with ONNX Runtime Web; this repository does not provide a hosted prediction API.
Files
| File | Role |
|---|---|
| clinical-28.onnx | Saved clinical text encoder used by the 28-day research candidate |
| clinical-90.onnx | Saved clinical text encoder used by the 90-day research candidate |
| radgraph-encoder.onnx | RadGraph-XL transformer component |
| radgraph-spans.onnx | RadGraph-XL entity/span and mention components |
| radgraph-relations.onnx | RadGraph-XL relation component |
The website separately includes the small risk networks, tokenizers, preprocessing, clinical-event construction and explanation logic. Downloading only these five files is not a complete standalone predictor. Exact byte sizes and SHA-256 values are listed in model-manifest.json.
The exports preserve the project's saved float32 parameters without quantization or distillation. The two clinical encoders contain the project's saved encoder states rather than unmodified upstream ClinicalBERT weights.
Intended use and endpoint
Research software for examining cumulative incidence of death after prediction time T0 and before liver transplantation, treating transplantation as a competing event. This endpoint is not 90-day all-cause mortality. The 28-day and 90-day candidates are distinct models.
The current report-processing path supports English reports. This repository contains no patient records or original clinical datasets. It is a research prototype and does not establish clinical utility or replace medical assessment. Model explanations describe contributions to a prediction, not causal or treatment effects.
Origins and notices
- ClinicalBERT base: https://huggingface.co/emilyalsentzer/Bio_ClinicalBERT
- RadGraph-XL source weights: https://huggingface.co/StanfordAIMI/RRG_scorers
- RadGraph implementation: https://github.com/Stanford-AIMI/radgraph
- RadGraph transformer/tokenizer: https://huggingface.co/microsoft/BiomedVLP-CXR-BERT-general
See THIRD_PARTY_NOTICES.md and LICENSE-RadGraph.txt for attribution. Upstream component licenses remain applicable; this card does not assign a new blanket license to all model components or grant access to underlying clinical datasets.