Learning center

BIDMC respiration: candidates, annotation agreement and event tables

Follow a respiration recording through cycle detection, annotation comparison and result export. Event agreement depends on the reference annotator and matching window; it is not clinical validation.

Follow a respiration recording through cycle detection, annotation comparison and result export. Event agreement depends on the reference annotator and matching window; it is not clinical validation.

1. Keep the WFDB record and annotation roles separate

The PhysioNet BIDMC PPG and Respiration Dataset example uses bidmc01.hea, its companion bidmc01.dat, and bidmc01.breath. Keep their original names and relationships. The header defines signal layout and sampling; the breath file supplies comparison events. The waveform contains 60,001 samples, five channels and 125 Hz sampling over 480 seconds.

Load the WFDB record and inspect its timebase, channel names and types before selecting Respiration Cycle & Variability. Assign Primary respiratory waveform to the measured RESP, channel, whose label includes punctuation in this source. PLETH is a different measured signal and is not an interchangeable respiratory waveform. The optional reference respiratory waveform is also distinct from an annotation-event layer.

The recorded configuration declares Sensor modality: impedance, Analysis polarity: auto, Inspiration direction: unknown, Cycle measurement layer: detected_candidates and output disposition: proposal. Retain those meanings. An algorithmically selected sign is not evidence of a physiologically identified inspiration direction, and impedance amplitude is not calibrated airflow, tidal volume or ventilation.

2. Preserve preparation and candidate settings

The recorded run uses a third-order respiratory-band filter with limits 0.05–1.0 Hz, minimum/maximum candidate rates of 4 and 60 breaths/minute, and prominence scale 0.50. It keeps missing-data and timebase audits with segment-safe preparation. Review any missing-run ledger, interpolation setting and prepared trace before accepting a candidate event. The waveform preview can be bounded for display; the input identity and complete candidate ledger define the analytical scope.

Choose the source breath label explicitly. The recorded GUI run selected ann1; a separate comparison evaluated ann2. If multiple source labels exist, do not merge them into one reference. Preserve the selected label, annotation count, matching tolerance and run parameters with every comparison.

The result registers 20 workflow sections, including prerequisite, QC and evidence sections. That is not a claim that all 20 have an unqualified computed status. The saved route ledger includes degraded sections for unreviewed cycles/unknown phase direction and an unavailable optional reference-channel comparison. Inspect each status and reason rather than treating missing output as zero or calling every section validated.

3. Read cycle timing without changing its meaning

The recorded example contains 169 complete, unreviewed cycle candidates. The ledger retains start, peak and stop sample indices and times, duration, candidate rate, rising/falling durations, filtered amplitude and area, and the measurement source. Complete-cycle duration is stop time minus start time; candidate breaths/minute is 60 divided by that duration.

For the first saved candidate, start is sample 66 at 0.528 seconds, peak is sample 177 at 1.416 seconds, and stop is sample 347 at 2.776 seconds. Duration is 2.248 seconds and rate is approximately 26.6904 breaths/minute. Check these values against the measured time axis rather than counting pixels in a figure. Rising and falling waveform intervals remain descriptive; unknown inspiration direction must not become an inspiratory/expiratory conclusion.

The reviewed controls also support a source-bound candidate-review ledger. A review must match the recording fingerprint and candidate digest; stale decisions are refused or cleared after an input change. The recorded example did not promote its detected candidates into clinically adjudicated breaths. Neither a review status nor “proposal” establishes diagnosis or physiological truth.

4. Compare the two annotators independently

The recorded event protocol uses a 0.50-second tolerance, one-to-one and order-preserving matching, maximizing matches and then minimizing total absolute timing error. Compare candidate peak times to the selected source event layer. The annotations do not establish a universal inspiratory or expiratory phase boundary.

LayerReference / matched / missed / extraRecorded interpretation
ann1165 / 165 / 0 / 4Sensitivity 1.0; PPV approximately 0.976331; F1 approximately 0.988024; median matched absolute timing error 40 ms.
ann2170 / 169 / 1 / 0Sensitivity approximately 0.994118; PPV 1.0; F1 approximately 0.997050; median matched absolute timing error 64 ms.

Timing error describes matched events only. The earlier 100 ms protocol on the reference-agreement page is a different tolerance; these figures are not a before/after improvement claim. This is one of 53 dataset records. It does not establish full-corpus or subgroup performance, apnea detection, clinical validity or calibration.

5. Save the ledger, route status and comparison context

  1. Review the measured/prepared trace, cycle boundaries, rate trend, quality flags and source-event agreement together. A visually smooth trace does not certify a cycle detector.
  2. Export the native result bundle. The recorded manifest identifies 61 artifacts plus the manifest itself: metrics, events, series CSV tables, figures and JSON context. Retain the input receipt, parameter manifest, recording fingerprint and selected source label.
  3. Open series_respiratory_cycle_candidates.csv and verify all 169 rows, durations and rates. Open the source-breath selection contract to confirm ann1, 165 selected events and no merged labels for this exported run. The independent ann2 benchmark is separate evidence, not an annotation layer hidden in this same export.
  4. Inspect series_route_status.csv and its reason codes. A bounded preparation preview has 15,001 rows in this recording; it does not replace the 60,001-sample WFDB source. Distinguish complete analytical ledgers from explicitly bounded visual previews.
  5. Check exported files against the native manifest’s byte counts and SHA-256 values, and open a PNG outside the viewer. Preserve the original source and settings for rerunning. This evidence does not establish a reloadable analysis session or current hosted download acceptance.

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

← Back to the learning center