Learning center

Biosignal annotations: refused loads, retained results and same-source recovery

Check annotation times and input identity, distinguish retained results from a new Run, and preserve the limits of one synthetic reference comparison.

A failed annotation load can leave a valid earlier result on screen. In Maternal/Fetal Signal Separation, check the active source, result identity and saved bundle before interpreting a comparison or exporting it.

This lesson follows one retained local refusal and recovery check using a synthetic two-channel recording. It explains annotation input and result identity; it does not establish fetal detection or separation performance on physiological recordings. Start with the Biosignal product path and calibration and unit contracts.

1. Keep waveform, metadata and annotations together

The checked fixture contains 385 samples at 128 Hz, with channels named maternal and abdominal. Its physiological-recording table, JSON metadata and six-row event table share the same filename prefix. The labels and declared mV units describe this synthetic fixture; they do not make it a clinical ECG or a validated maternal/fetal mixture.

InputChecked contractBefore loading
Waveform and metadataThe fixture uses matching _physio.tsv.gz and _physio.json files. Metadata declares 128 Hz, start time 0, two channel names and mV.Keep the actual sampling rate, channel names and unit metadata with the source; do not infer them from a plot.
Event filenameThe loader looks beside a BIDS-named waveform for the same prefix plus _events.tsv.A missing event table yields no annotations. Successful waveform loading alone does not establish reference availability.
Event timesonset and duration must be finite numeric values in this loader; duration must be zero or positive. The six retained durations are zero.Use seconds, not sample indices or milliseconds. Keep onset and duration distinct; the refusal below concerns duration.
Reference identityThe fixture has trial_type equal to fetal_qrs and channel equal to abdominal. Annotation metadata retains the event-file hash.Check the intended event label and time reference. A channel name alone does not assign a fetal-reference role.

The BIDS event specification defines onset and duration in seconds, permits negative onsets for events before the first stored data point, and permits zero-duration events. Its missing-value conventions are broader than this loader’s finite-number requirement. This example did not exercise negative onset or missing times and does not establish general BIDS conformance.

2. Establish the active source before a result

In an authorized local workflow, open Maternal/Fetal Signal Separation and use its local Load control for the waveform. Check that loading completes, that the source field identifies the intended file, and that both channels and all six reference events belong to that input. Do not interpret a populated result pane as confirmation that a replacement file loaded.

The retained comparison selected Recognized fetal-reference comparison and the claim-boundary route. Its tolerance was 0.0625 s; this is the recorded setting, not a recommended clinical threshold or the default for every session. Channel roles and all other parameters remain in the saved parameter manifest. No separate Validate or Apply action was exercised. Loading and running are separate actions.

3. Read a refused load as retained prior data

The checked invalid copy changes only the first event’s duration from 0 to −0.0078125 s. The loader refuses it before replacing the active recording. The status reads Local recording load failed; prior data remain active. The original source field, six annotations, earlier result and viewer remain in place.

Saved stateActive source and resultWhat Export means
Valid baseline exportedOriginal valid recording; first result and viewer share a run ID.Exports the first result. Its bundle contains 31 artifacts and a separate manifest.
Invalid annotation refusedSame recording, original events and first result. The file field is restored.The button remains enabled for the original result. No Export was invoked during this refusal check.
Original valid source reloadedThe same input identity is loaded into a new recording object. The earlier result remains until a fresh Run.Still refers to the earlier result before that Run.
Recovery Run completed and exportedSame valid source; a new result and viewer share a distinct run ID.Disabled while the Run is pending, then enabled for the new result. A second bundle is added; the first is unchanged.

An enabled Export button is therefore not proof that the most recently attempted input was analyzed. Read the load status and source identity first. A retained plot is evidence of the earlier result until a new Run completes.

4. Recover without losing the earlier bundle

  1. Preserve the earlier export and the refused input. Determine which annotation field caused the refusal.
  2. Load the intended valid source and confirm its identity. The recorded recovery reloaded the identical original valid source; it did not test switching to a different dataset.
  3. Check channel selection, reference annotations and analysis parameters, then Run. Wait for completion before exporting.
  4. Export the new result and retain both run directories. Compare the new manifest, input receipt and result review with the intended source and settings.

Use the selected-recording fingerprint consistently across the export manifest, input receipt and result review. The retained loaded-recording fingerprint differs from the selected-channel result fingerprint; comparing those two fields directly would be misleading. Also retain the waveform hash and annotation-file hash: identical waveform bytes alone do not describe the event table used.

5. Read the one synthetic comparison within its limits

The earlier independently checked comparison has six reference events and six candidate events. At its recorded tolerance, five pairs match, one reference is missed and one candidate is extra. The saved signed timing errors for the five matches are −0.015625 s, +0.015625 s, and three zeros. Sensitivity, positive predictive value and F1 are each 5/6; the median absolute error is zero. These are retained results from that one fixture, not a fresh matching analysis or an estimate of clinical accuracy.

Read all seven rows of the comparison table. The missed reference has no matched candidate time; the extra candidate has no reference time. Empty unmatched cells are not zero-second errors. The timing figure shows five matched pairs and cannot substitute for those complete rows.

Complete saved comparison

These tables transcribe the earlier accepted comparison, not another recovery Run. The comparison CSV preserves seven columns and seven rows in the order shown below. On a narrow screen, focus the table and use the arrow keys to scroll across all columns. A dash marks an empty exported cell, not zero. Indices are positions in the sorted event lists, not waveform sample indices; their .0 suffix is retained from the CSV.

statusreference_indexreference_time_secondsmatchedcandidate_indexcandidate_time_secondssigned_error_seconds
TP0.00.265625True0.00.25-0.015625
TP1.00.734375True1.00.750.015625
TP2.01.25True2.01.250.0
TP3.01.75True3.01.750.0
TP4.02.25True4.02.250.0
FN5.02.9375False———
FP——False5.02.75—

The signed error is candidate time minus reference time, in seconds. A negative value means that the candidate is earlier. The missed reference at 2.9375 s and extra candidate at 2.75 s have no assigned partner and no signed error. They remain separate rows; the figure's five matched positions omit them.

All nine comparison metrics

The metrics below are all rows whose CSV scope is fetal_reference. The full metrics file has sixteen rows; the other seven describe input, candidate, template and preparation observations. The CSV unit cells for these nine rows are empty; the timing measure's name identifies seconds. Sensitivity, positive predictive value and F1 are proportions, saved as the floating-point value below.

Saved measureSaved value
reference_count6
candidate_count6
true_positive5
false_positive1
false_negative1
sensitivity0.8333333333333334
positive_predictive_value0.8333333333333334
f10.8333333333333334
median_absolute_error_seconds0.0

The retained independent reference expresses each proportion as 5/6. The CSV's decimal digits reflect floating-point serialization. A median absolute error of zero describes only the five matched pairs: three errors are zero, while two are nonzero. It does not remove the one missed reference or extra candidate, and it does not establish detector accuracy on physiological recordings.

In the earlier recovery bundles, all six fetal candidates remain unreviewed. Their route is marked degraded_dependency, and the exact result review remains unreviewed with proposal disposition. A computed comparison and a true prerequisites_met field do not supply candidate review or scientific approval. The two selected routes require fourteen routes in the dependency graph; this is not fourteen independent method validations.

6. Separate candidate annotations from result review

A separate retained workflow used the same synthetic fixture to follow one staged candidate rejection. It loaded the fixture once, ran it to produce R1, staged one decision, then ran again without reloading to produce R2. Those two results belong to this separate review check; they are not the earlier recovery exports.

  1. Start from a completed result and identify the intended candidate in Source-bound fetal-QRS candidate review. The recorded example selected fetal_candidate_0000001, at 0.25 s.
  2. Record the decision, reviewer role and reason. This test selected reject_fetal_qrs, with role synthetic fixture metadata probe and reason Test-only annotation propagation; not a physiological review. These describe the test, not a judgment about a physiological event.
  3. Use Stage decision. In the retained state, one decision became pending while the existing R1 result, candidate table and run generation stayed unchanged. Staging did not apply the annotation to that result or start another analysis.
  4. Run again with the same source and settings to bind and apply the staged annotation. The recorded R2 has a new run ID, one review-log entry and the corresponding candidate/event annotation. Verify those result fields together; a pending-status message alone does not establish application.

The source associates a staged decision with the selected-recording fingerprint and base candidate-table digest. In this one workflow, all 385 sample pairs, time values, six source events and metadata stayed unchanged. The selected fingerprint was the same in R1 and R2, while their run IDs differed. No new calculation was performed for this lesson.

Read every candidate decision

The table shows all six candidate rows before and after this separate review Run. On a narrow screen, focus the table and use the arrow keys to reveal all columns. The full sample indices, candidate fields, reviewer role, reason and event annotations remain in the downloadable evidence.

Candidate IDOnset (s)R1 decisionR2 decision
fetal_candidate_00000010.25unreviewedreject_fetal_qrs
fetal_candidate_00000020.75unreviewedunreviewed
fetal_candidate_00000031.25unreviewedunreviewed
fetal_candidate_00000041.75unreviewedunreviewed
fetal_candidate_00000052.25unreviewedunreviewed
fetal_candidate_00000062.75unreviewedunreviewed

The rejection changes review metadata. All six base candidates remain, and the automatic comparison still uses the original detected peak positions. Both saved results have the same seven comparison rows and nine comparison metrics: five matches, one missed reference and one extra candidate. The source does not remove a rejected candidate from that automatic comparison. An annotation count of one is not a revised detector score or a validated physiological decision.

Keep the exact result review separate

A candidate annotation and an exact result review serve different purposes. In both saved results, the exact result review was not supplied, its decision stayed unreviewed, and the output disposition stayed proposal. The candidate route retained degraded_dependency; the five scientific-claim flags remained false.

Recorded fieldR1R2
Exact review suppliedfalsefalse
Exact review decisionunreviewedunreviewed
Output dispositionproposalproposal
Review-document SHA256ffa9bcc467b74c6f958f1db9152be2b113b4da1dcd80db4899454d4da40535a90f3fcda1a41da4113162edc8033f71f35efe25cd5029159e35b1422e00ba754b

The document hash changes because the document describes the exact result. Both complete nine-string documents were retained and their serialized bytes hashed; each matches its own result summary and in-memory manifest row. The component digests inside those documents remain opaque here. Their presence does not independently verify the scientific artifacts they name, and these manifest observations are not new native exports.

The earlier attempt failed because its checker expected the derived document hash to stay unchanged; its receipt also exceeded the declared receipt-size cap. That failed attempt remains preserved. A later harness recorded and checked both actual documents, with unchanged product source; the earlier failure was not relabeled as a pass.

The successful retained check invoked native controls programmatically and saved no exports or screenshots. Its Run control was recorded outside the visible window bounds, so native visual usability remains unreviewed. One pending decision was still observed after R2; no general persistence, clearing, revision, duplicate-decision or different-source guarantee follows. The earlier recovery bundles do not establish export of this R2 review.

A saved-only source review dated 11 November 2026 matches the application, toolbox helper, review foundation, window handler and data-model files to this recorded workflow. The shared UI helper differs, as already documented for the recovery guide. This section adds no whole-GUI, detector, separation, real-recording, clinical or release acceptance.

7. Save the complete recovery bundle and its meaning

Each of the two recovery-check bundles contains 27 CSV files, three JSON artifacts and one figure, plus a separate manifest: 32 files per run. The earlier accepted comparison is reused for setup and recovery. These are two saved results of the same fixture, not two independent datasets.

Saved materialRead it forLimit to retain
Manifest and input receiptRun ID, selected-recording fingerprint, channel selection, waveform identity, parameters and artifact hashes.Use the matching result identity; do not substitute a loaded-recording fingerprint or a filename alone.
Event and candidate CSVsAll six source events, six candidates, candidate decisions and the review log.An empty review log and unreviewed decisions remain unreviewed; export does not supply approval.
Reference comparison and route-status CSVsAll seven comparison rows, signed errors, unmatched rows and dependency status.The five-pair figure is a view of the matched rows. Required routes and computed dependencies are not independent validation.
Result review, parameter manifest and remaining CSVsExact result disposition, full saved settings and supporting preparation, residual, governance and resource tables.The complete files were preserved. Their presence does not independently validate residual amplitudes, filtering or separation.

The saved bundle review checks all 62 artifact identities across the two runs, both manifests and the structure of all 54 CSVs. Its table-structure checks cover 2,700 data rows and 10,010 cells; those counts describe completeness, not newly verified scientific values. The prior matching reference is reused without another calculation.

8. Keep the source and evidence boundaries visible

Five named current files—the application, shared toolbox helper, loader, data model and window handler—match the accepted payload and recorded recovery sources. The current shared UI helper differs from that payload, so these receipts do not establish a new whole-GUI review of current source.

The recovery evidence covers one negative duration and reloading the same valid source. It does not establish different-source stale-result handling, running cancellation, duplicate-event behavior, a malformed-input suite or a security boundary. The expected error message was captured through a messagebox callback; actual modal rendering was not reviewed.

The fixture contains no true maternal component mixed into the abdominal signal. Three maternal candidate centers, three complete template windows and finite residual values were prerequisite observations, not independent evidence for extraction accuracy, ECG morphology, filtering or residual amplitudes. No real-recording, patient, clinical, hosted-service, whole-app or release acceptance follows from this lesson.

9. Evidence and related learning

Download the scoped source and saved-evidence ledger. The additive comparison detail retains all seven comparison columns and all sixteen metrics rows from that earlier native bundle. A separate supplement records the staged-review workflow, both complete result documents and its historical failure; it contains no new native export. It identifies the accepted comparison and recovery reviews, exact sources, four saved states, fixture mutation and export identities. No scientific Run, native export or recording was repeated for this guide.

Continue with calibration, units and review disposition, the separate BIDMC respiration example, or all written guides. The BIDMC example has its own input and annotation scope; it does not expand this synthetic maternal/fetal check.

Worked examples describe particular inputs, methods and recorded results. Check current availability and input requirements before planning an evaluation.

← Back to the learning center