JakJak504's picture shoupoudou's picture
Add five-fold ancestry ensemble, reference tables and model card (#2)
961e9c7
|
Raw History Blame Contribute Delete
2.8 kB
metadata
tags:
  - ecg
  - electrocardiogram
  - genetics
  - ancestry
license: other
license_name: to-be-determined
extra_gated_prompt: >-
  These weights are derived from participant ECG recordings held under
  data-access agreements. Access is granted per project. State your institution
  and the approved use.

DeepECG-SSL ancestry -- five-fold ensemble

Predicts P(non-European genetic ancestry) from a 12-lead ECG. One logit through a sigmoid: p(European) = 1 - score. Five disjoint cross-validation folds, averaged.

Work in progress and unpublished. Quantitative performance, validation results and the cohorts involved are deliberately not described here; they are reported in the accompanying manuscript once it is published. Request access if you need to evaluate the model before then.

Layout

raw/fold_{1..5}/bs_128/checkpoint_best.pt   the five fine-tuned heads
reference/                                  frozen tables the pipeline compares against
MODEL_PROVENANCE.txt                        md5 of every checkpoint, and what produced it
CHECKSUMS.txt                               md5 + sha256 of every file here

The backbone is not in this repository. It is base_ssl.pt in the public, pinned heartwise/wcr_77_classes snapshot -- verified byte-identical (sha256 39985e904bfcc4e3fa58f3e631d84598c7228bfa821a3279109dc2e05f54383c) to the backbone these folds were fine-tuned from. Repinning that repository means re-checking that hash: a different backbone does not fail, it silently changes every score.

Use

pip install -r requirements.ancestry.txt          # DeepECG_Docker, branch DeepECG-Ancestry
export MODELS_ROOT=/path/to/this/snapshot
export REFERENCE_DIR=/path/to/this/snapshot/reference
export SSL_CKPT=/path/to/base_ssl.pt              # or let the manifest fetch it
python -m ancestry weights                        # confirms what resolved from where

Input is 12-lead, 2500 samples at 250 Hz, millivolts, lead-major. The loader refuses microvolts and raw ADC counts, because wrong units do not raise -- they give meaningless scores.

Limits

  • The label is binary European / non-European from genome-inferred ancestry. It is not self-declared race and must not be used as one.
  • Calibration depends on cohort prevalence. Recalibrate before any thresholded use.
  • Accuracy is not uniform across non-European groups.
  • Site effects can dominate in cohorts recorded on different hardware.
  • Not a clinical device. No outcome, diagnosis or treatment decision should rest on it.