physiotrace

Does your task head rely on how the data were recorded?

physiotrace audits a linear task head on a frozen encoder for reliance on acquisition provenance — device, site, reference, montage, scanner, batch — and certifies a remedy only when it is specific and safe. It works on embeddings, so any frozen model (EEG, ECG, imaging, omics) can be audited without retraining.

Two common measurements do not answer this question. Shifted-transfer benchmarks can fail for unrelated reasons, and a provenance probe shows only that provenance is decodable, not that a head uses it. PhysioTRACE tests the head's behavior directly:

Every audit ends in one of three verdicts: no reliance, reliance – remedy certified, or reliance – remedy not certified.

New here? Start with the tutorial: bring your own embeddings, choose the settings, read the report, and avoid the mistakes that make an audit misleading.

Install

pip install physiotrace
pip install "physiotrace[plot]"   # with matplotlib, for the tutorial figures

Python 3.10–3.13. Tested on every push from the oldest supported dependencies (numpy 1.24, scipy 1.10, scikit-learn 1.3, pandas 2.0) to the latest.

Quickstart

Observational data (for example device or site)

import physiotrace as pt

train = pt.Split(embeddings, y, u, clusters=patient_id)   # u: e.g. device A / device B
test = pt.Split(test_embeddings, test_y, test_u, clusters=test_patient_id)

report = pt.audit_observational(train, test, epsilon=0.01, delta=0.01)
print(report)
report.verdict        # pt.Verdict.REMEDY_CERTIFIED / REMEDY_NOT_CERTIFIED / NO_RELIANCE
report.to_dict()      # JSON-serializable summary

PhysioTRACE audit (observational)
  verdict: reliance, remedy certified
  Recover   conditional_accessibility_auroc = 0.9946
  Stress    G = 0.1384 [0.1228, 0.1539]  (epsilon = 0.01; 95% interval)
  Intervene G after rank-1 removal = -0.0000 [-0.0013, 0.0012]  (100% reduction)
  Verify    specific: yes (beats 100/100 random controls)
            safe: yes (control-head AUROC, independent regime change -0.0000 [-0.0004, 0.0003], delta = 0.01)

(Output of examples/quickstart_synthetic.py, which also shows the other two verdicts, a cross-fitted audit, and a paired audit.)

y and u can hold any two labels each (strings or numbers). clusters identifies who each record belongs to (patient, subject); it is the bootstrap unit, and a cluster on both sides of the split is rejected. epsilon is the smallest G that counts as reliance and delta the tolerated utility loss; declare both, and the localization rank, before looking at test results.

The exposed head is fitted on the natural association and the control head with independence weights. Provenance is localized conditionally on the target, so removal does not by default discard population signal that sources differ in.

To cross-fit over folds instead, as in the paper's PTB-XL audit:

report = pt.audit_observational_crossfit(embeddings, y, u, folds=folds, clusters=patient_id)

Checks

Every report lists validity checks under report.findings and in print(report): too few records in a (y, u) cell, a small effective sample or few effective clusters after reweighting, a very large weight, a stressed association too weak to test, a head that did not converge, missing clusters, and too few bootstrap draws. Checks never change the verdict; those of severity warning are also raised as pt.PhysioTraceWarning. Thresholds live in pt.Gates, and pt.Gates.paper() gives the stricter gates of the paper:

report = pt.audit_observational(train, test, gates=pt.Gates.paper())

Common input mistakes fail early with a specific message: a target with more than two classes, a provenance factor with more than two groups, a patient in two folds or on both sides of the split, or a (y, u) combination with no records.

Paired views (for example EEG references of the same record)

train = pt.PairedSplit(embeddings, y, views, records)   # one row per (record, view)
report = pt.audit_paired(train, test, val=val, p_shortcut=0.6, delta=0.02)

The association is induced by a designed coupling at fixed marginals; each view serves in turn as the shortcut and G averages over them. Provenance is localized from within-record view contrasts without task labels. report.details adds decision flip rates across views and the same-record logit contrast.

Building blocks

Every step is also available on its own:

pt.lpi(train_h, train_u, val_h, val_u, test_h, test_u, test_clusters=subject)  # Recover
tables = pt.observational_regimes(y_train, u_train)                             # Stress regimes
w = pt.cell_weights(y_test, u_test, tables["reversed"])
pt.weighted_auroc(y_test, scores, w)
subspace, probe = pt.conditional_direction(z_train, y_train, u_train)           # Intervene
subspace.remove(z_test)
pt.random_subspaces(subspace.whitening, rank=1, n=100, seed=0)                  # Verify controls

physiotrace.synthetic generates embeddings with a known shortcut that reproduce each of the three verdicts on purpose.

Interpreting the result

• A certified remedy supports the evaluated association and the linear intervention, not safety under every deployment shift.
• LPI and accessibility never yield a verdict on their own: a head can ignore what a probe decodes.
• Declare epsilon, delta, the rank, and the number of random controls before looking at test results.
• Keep every identity (patient, subject, record) inside one fold and pass it as clusters; resampling windows or trials instead overstates precision.

Validation against the paper

validation/ reproduces the published numbers of the PhysioTRACE paper from its saved artifacts:

The paired NMT run whitened with an SVD; pass whitening="svd" to draw the identical random controls. Both methods give the same projections.

Status

Version 0.1. Binary targets with binary provenance (observational, target-conditional localization of any rank; rank 1 as in the paper) or K paired views.

License

MIT.