Reference

Analysis catalogue

What Oxynet measures in a CPET, and how much weight each measurement will bear. Every entry states what it needs, what it returns, the evidence behind it and where it stops.

API v1 · docs 1.0 · updated 2026-09-22

Read with every row

Nothing here is calibrated against clinical outcomes, and no output carries a confidence score. What has been measured is agreement with expert labelling, which is a different claim. Oxynet is research software, not a medical device, and its output is not a diagnosis.

Production

Answering on the live API, with the evaluation that let each model ship stored beside it.

Ventilatory thresholds (VT1, VT2)

Production

The first and second ventilatory thresholds, located by a convolutional network over the breath series.

Needs: Time, V̇O₂, V̇CO₂, V̇E, PetO₂, PetCO₂. The channels a given model needs are checked before it runs.
Returns: vt1_time_s, vt2_time_s, vt1_vo2_ml_min, vt2_vo2_ml_min, with the model and version that produced them.
Evidence: Agreement with expert labelling across cohorts that differ in population, protocol, ergometer and metabolimeter, in peer-reviewed work. The longest record of any capability.
Limits: Agreement with expert raters, not calibration against outcomes: no confidence score is returned. A missing gas channel is refused, not substituted.
Call: POST /v1/cpet/{id}/analyze {"analyses": ["vt"]}

Exercise intensity domains

Production

Moderate, heavy and severe domain probabilities per second, from the same network as the thresholds.

Needs: As for thresholds.
Returns: Three class probabilities per second; the thresholds are where they cross.
Evidence: As for thresholds: the same model.
Limits: Not returned by /v1/analyze, which returns the thresholds only.
Where: Shown in the web application at app.oxynet.net.

Substrate use and FATMAX

Production

Fat and carbohydrate oxidation by indirect calorimetry, FATMAX, the crossover to carbohydrate, and gross efficiency.

Needs: V̇O₂ and V̇CO₂. A work-rate channel adds a wattage axis and gross efficiency; without one it runs against %V̇O₂peak.
Returns: Windowed oxidation rates, FATMAX (a fitted turning point, not a measured sample), crossover, efficiency, and flags.
Evidence: Standard stoichiometry (Frayn 1983; Jeukendrup and Wallis 2001), with the equations returned in the response.
Limits: Above RER 1.0 the fat rate is an upper bound, not a measurement. A ramp never reaches steady state, so FATMAX reads high and is not comparable with a graded test. Both are flagged.
Call: POST /v1/cpet/{id}/analyze {"analyses": ["substrate"]} (or the substrate_summary metric)

Derived CPET quantities

Production

Peak values, O₂ pulse, the V̇E/V̇CO₂ slope profile, rest / warm-up / exercise / recovery phases, and classical landmarks.

Needs: Per metric; GET /v1/metrics states each one’s required channels and minimum duration.
Returns: vo2max, vemax, rermax, o2_pulse, ve_vco2_slope (over 25/50/75/100 % of exercise, with intervals), phases, landmarks, substrate_summary.
Evidence: Deterministic computations on the recording’s own signal.
Limits: Landmarks (V-slope, respiratory compensation, V̇E/V̇CO₂ nadir, PetCO₂ peak) corroborate and never override a model. Phase boundaries carry an error of tens of seconds.
Call: POST /v1/cpet/{id}/compute {"metrics": [...]}

Signal integrity

Production

Whether the recording can support a measurement at all: cross-channel agreement of the gas analysis, the band the sampling can resolve, and clock repair.

Needs: Any parsed recording.
Returns: gas_quality, sampling_adequacy, clock_integrity; channel presence, coverage and parser flags in the upload summary.
Evidence: Physiological consistency checks: CO₂ output against ventilation, end-tidal O₂ against CO₂.
Limits: An absent check is not a passed one: n_checks says how many were runnable.
Call: GET /v1/cpet/{id} and POST /v1/cpet/{id}/compute

Vendor format parsing

Production

Reads the export as the cart wrote it, with each format’s units, clock resets and decimal conventions.

Needs: The unchanged export file.
Returns: A cpet_id, and a summary of what was parsed and what is absent, with reasons.
Evidence: Detection is scored from content and refuses when no candidate is clearly best.
Limits: Some formats rest on few example files; the formats page says which. A different software version may not parse.
Call: POST /v1/cpet

Research

Running, and not yet shown to transport to a new population. Treat outputs as research findings.

Oscillatory ventilation (EOV) and the ventilatory control loop

Research

A continuous wavelet detector for oscillatory breathing, its grade and burden, and the loop gain of the chemoreflex behind it.

Needs: V̇E and PetCO₂ (at least 80 % coverage) over at least 128 s. Breathing frequency separates rate from volume drive.
Returns: Grade and burden, events with period and amplitude, CO₂ coupling, the loop-gain margin, gas quality.
Evidence: Developed on a single heart-failure cohort of 44 patients.
Limits: Beta. Transportability to any other population is untested. Treat every output as a research finding.
Call: POST /v1/cpet/{id}/analyze {"analyses": ["eov"]}

Alternative threshold engine

Research

A second threshold model offered as a second opinion.

Needs: As for thresholds.
Returns: VT1 and VT2.
Evidence: Built and checked on a single group of 100 clinical tests.
Limits: Not tested anywhere else. Deliberately not on the API or MCP.
Where: In the web application only, marked beta.

Self-supervised pretraining

Research

Pretraining on unlabelled recordings.

Needs: Not applicable.
Returns: None exposed.
Evidence: In development.
Limits: In no promoted model, and not reachable from any interface.

How a model reaches production

A threshold model is promoted only after evaluation on a cohort it never saw, on both axes, clearing all four of: V̇O₂ bias ≤ 120 mL/min, V̇O₂ correlation ≥ 0.80, time bias ≤ 60 s, time correlation ≥ 0.80. Every promoted model ships with the evaluation that let it through, and its name and version are in the provenance of every result it produces.