dRFI calibrated 50-model ensemble
This model returns a calibrated probability of a dRFI event within ten years for each patient.
Inputs
Feature H5s must use PolarisFMv3-small
revision 0dc4c3534027bfb7a24ec8a8d965f06f9aecbb63, cat-max pooling,
and 224-pixel, 224-µm tiles.
SAI Engine checks the extractor repository and revision recorded in each
feature H5 against the model config.
The clinical CSV uses these columns: age, t_stage, n_stage,
histo_grade, er_status, pr_status, her2_status, and
histology_subtype. Accepted histology subtypes are ductal, lobular,
mixed, and other. Leave a missing categorical value empty rather
than recording it as other. See config.json for the other accepted
categories.
Run inference
Use SAI Engine v2.3.0 or later and authenticate to this private repository.
sai-predict \
--model-url https://huggingface.co/StratifAI/dRFI-ensemble-calibrated \
--feature-dir ./features \
--clini ./clinical.csv \
--slide-csv ./slides.csv \
--patient-col PatientID \
--output ./predictions.json
The slide CSV maps patient IDs to feature H5 paths. If each H5 represents
one patient, omit --slide-csv; SAI Engine uses the filename stem as
the patient ID.
Output
predictions.json maps patient IDs to calibrated ten-year event-risk
probabilities:
{
"PAT-001": {"event_risk": 0.08}
}
The value is a probability from 0 to 1, not a percentage.
- Downloads last month
- -