psgscoring

Open-source AASM-compliant respiratory scoring for polysomnography.

Paper

Rombaut B, Rombaut B, Rombaut C, et al. Graded evidence in place of thresholds: an open-source, AASM-compliant method for respiratory event detection in polysomnography. Manuscript in preparation, 2026.

Technical details (signal processing chain, classification logic, bias corrections): Technical Reference (Online Supplement)

What this library does

psgscoring detects and classifies respiratory events (apneas, hypopneas, RERAs) in polysomnography recordings following AASM rules. It extends YASA (Vallat & Walker, eLife 2021) from sleep staging into a complete clinical respiratory scoring pipeline.

Four things that distinguish this library:

1. Graded evidence instead of a threshold cascade — the AASM Rule 1A conjunction is evaluated as a product of graded terms (flow reduction, duration, desaturation-or-arousal, breath template) rather than a chain of yes/no cuts. One operating point replaces three parameter combinations. On MESA (n=150, held out) this raises event-level agreement over the rule cascade by a median ΔF1 of +0.029 (p = 7·10⁻⁸) without costing anything on the AHI itself.
2. AHI confidence interval — every study is scored at three stringency levels (strict/standard/sensitive), yielding a per-study robustness grade (A/B/C) rather than a single AHI number.
3. Clinical auditability — every event carries a confidence score, the rule that admitted it, the graded terms behind that decision, and the counters of every correction that touched it. Rejected candidates are kept with their rejection reason rather than discarded, so a reviewing physician can see what was not scored and why.
4. Measured, not asserted — every behavioural change ships with the measurement that motivated it and a decision rule fixed before the sweep. CHANGELOG.md carries the numbers, including the ones that argued against the change. Validation runs on PSG-IPA (5 recordings, 12 scorers each) and MESA/NSRR (held out, nothing tuned on it).

Release policy

Releases are measurement points, not milestones. Several may be cut on the same day: every behavioural change is validated against the golden harness and at least one cohort before it ships, and the version is what makes that validation citable. A high version number reflects how often the work is measured, not instability.

The stability guarantee lives in the profiles, not in the version number:

• mesa_shhs and chicago_1999 are frozen. They reproduce published results (paper v31/v37 and the 1999 Chicago Criteria respectively) and are pinned against behavioural changes, including repairs that every other profile receives.
• All other profiles track the current best understanding of the AASM rules and may move between releases when a measurement justifies it. Each such change is behind a profile field whose default preserves the previous behaviour, is recorded in CHANGELOG.md with the measurement that motivated it, and is pinned by a test.

If you need scored values to stay identical across time — for a study, a regulatory submission, or a paper — pin the version (psgscoring==0.18.0) and record which profile you used. Do not rely on a profile name alone.

Installation

pip install psgscoring

Requirements: Python ≥3.9, numpy, scipy, mne. No GPU required.

Quick Start

import mne
from psgscoring import run_pneumo_analysis

# Load EDF and provide a hypnogram (e.g., from YASA)
raw = mne.io.read_raw_edf("recording.edf", preload=True)
hypnogram = ["W", "N1", "N2", "N2", "N3", ...]  # per 30-s epoch

# Run the full pipeline
results = run_pneumo_analysis(raw, hypnogram, scoring_profile="aasm_v3_rec")

# Access results
resp = results["respiratory"]["summary"]
print(f"AHI: {resp['ahi_total']}, Severity: {resp['severity']}")
print(f"Events: {resp['n_obstructive']} OA, {resp['n_hypopnea']} Hyp")

# AHI confidence interval
interval = results["ahi_interval"]
print(f"AHI interval: [{interval['strict']['ahi']}–{interval['sensitive']['ahi']}]")
print(f"Robustness: {interval['robustness_grade']}")

Scoring Profiles

list_profiles() enumerates the full registry (AASM v1/v2/v3, CMS/Medicare, Chicago 1999, and the NSRR dataset profile).

aasm_v3_breath — breath-graded hypopnea scoring (v0.13.0, opt-in)

run_pneumo_analysis(raw, hypnogram, scoring_profile="aasm_v3_breath")

A second hypopnea detector that takes the breath, not the sample, as the unit. The AASM speaks of peak signal excursions; this scores them directly. Five differences from the default detector:

1. Breath-level events. An event is a run of consecutive breaths below the threshold, so boundaries land on breath transitions and a recovery breath ends the run by itself — no smoothing, and the merge/split problem that depresses event-F1 largely disappears.

2. Two-pass patient calibration. Pass 1 finds the incontestable events; pass 2 measures the baseline from only the breathing that is not an event, and derives that patient's own SpO₂ delay by cross-correlation instead of assuming a fixed 30–45 s window.

3. Graded AASM predicates. The rule structure stays literally the Rule 1A conjunction — ≥30% reduction AND ≥10 s AND (≥3% desaturation OR arousal) — but each threshold gets a tolerance rather than being infinitely sharp.

4. Every event carries p_scored plus a criteria dict giving each predicate's contribution, so the audit trail says why, not just what.

   p_scored ranks events by how well they satisfy the AASM conjunction. It is not the probability that a scorer would mark the event: measured against the 12-scorer fractions on PSG-IPA (163 events) the correlation is only r = 0.194 and the level is +0.33 too high. Use it to order events, not as a likelihood.

5. One strictness axis (hypopnea_strictness, default 0.50) instead of three parameter combinations.

Scope is hypopneas only; apneas keep the existing detector.

Status — opt-in, not the default. On PSG-IPA (5 recordings, 12 scorers) the median event-F1 rises 0.343 → 0.434, the percentile within the inter-scorer distribution p6 → p17, and mean |ΔAHI| against the scorer median falls 1.84 → 0.29. On MESA the paired advantage replicates on two disjoint held-out samples of 50 (p = 0.0069 and p = 0.0016). But the two datasets disagree on the absolute level by ~16 AHI points, and until that is explained the existing clinical output stays the default. See docs/interim_conclusie_klinisch_gebruik.md.

aasm_v3_dual — apneas on both flow sensors (v0.14.0, opt-in)

run_pneumo_analysis(raw, hypnogram, scoring_profile="aasm_v3_dual")

The AASM assigns apneas to the oronasal thermistor and hypopneas to nasal pressure. Choosing one sensor means inheriting its blind spot: nasal pressure misses mouth breathing, the thermistor is too slow for short events. This profile detects apneas on both and merges them, so neither channel can veto the other.

Each merged apnea carries corroboration — both, thermistor_only or pressure_only. Nothing is discarded on that basis. Across 1785 apnea detections on MESA only 19% are seen by both sensors, so requiring corroboration would remove four events in five. Discarding can be licensed explicitly (corroborate_apnea_events(..., corroboration_licensed=True)); no profile does so.

meta["flow_channels"]["thermistor_check"] reports the envelope correlation between the two channels either way. Under this profile it is informational — the second sensor is additive and can only add events. It gates only the substitutive path, where a dead thermistor would delete them.

Why this exists. v0.13.2 made the thermistor the apnea sensor outright. On a montage where that channel carried no usable breathing signal, apneas went from 93 to 0 and the conclusion moved from moderate CSAS — human-confirmed — to mild SAS. That release was rolled back. aasm_v3_rec remains the default and is byte-identical.

aasm_v3_pressure — nasal-pressure reference (v0.14.1, opt-in)

run_pneumo_analysis(raw, hypnogram, scoring_profile="aasm_v3_pressure")

Apnea detection is not the only consumer of a flow signal. The AHI robustness sweep, baseline anchoring, the arousal/RERA coupling, Cheyne-Stokes detection and the ventilatory burden each take one — and each took the apnea channel. This profile scores apneas exactly as aasm_v3_rec does and points those five at the nasal pressure instead.

¹ where it passes the quality check; nasal pressure otherwise.

The AASM assigns quantitative flow assessment to nasal pressure and treats the oronasal thermistor as qualitative — able to show that flow stopped, not to grade how much of it is left. Reading a thermistor for the ventilatory burden therefore measures something the sensor does not claim to report: on one recording it read 20.4% instead of 42.6%, below the ≤25% reference, for a patient spending 94.7% of the night under 90% saturation.

Identical to aasm_v3_rec on any montage without a usable thermistor. Set flow_reference on any profile to get the same behaviour; meta["flow_channels"]["reference_sensor"] reports which channel was used.

Validation

PSG-IPA (PhysioNet): 5 recordings, 59 independent scorer sessions. Against the scorer median, aasm_v3_rec gives bias +1.69/h, MAE 1.76/h, Pearson r = 0.997, weighted κ = 0.839, severity concordance 4/5. Note the sample size: five recordings is enough to expose a defect, not enough to estimate a population.

MESA (NSRR, external cohort, n = 150): nothing is tuned on it. Against a reconstructed AASM-2015 reference, per profile:

Paired against aasm_v3_rec, breath-graded scoring raises event agreement by a median ΔF1 of +0.029 (95/150 recordings, p = 6.8·10⁻⁸); the probabilistic variant by +0.036 (p = 1.4·10⁻⁹).

These numbers moved substantially in 0.17.0, and the reason is worth stating plainly. The earlier figures on this cohort reported a bias of −11 to −15/h and were read as under-detection. Most of it was not: a signal-quality gate was rejecting effort channels by how the EDF declared its unit rather than by the signal in them, and it failed 52 of 52 MESA recordings. Repairing it halved the bias at identical F1, precision and recall on three of the five profiles — the events had been found all along and were being discarded in the accounting. See CHANGELOG.md 0.16.0 and docs/.

Bias corrections

The twelve below were the original systematic pass: six over-counting and six under-counting mechanisms, each identified and quantified separately. The list has grown since and is no longer the whole story — CHANGELOG.md carries the current state, with the measurement behind every entry. The most consequential recent one: a sensor-quality gate that judged how the EDF declared its unit rather than the sensor itself.

Since then (see CHANGELOG.md for the measurements):

Four of these are off, and that is a result, not an omission. Each was implemented, measured, and left disabled because the measurement did not support enabling it: the desaturation limit removed 8 events in 4303; breath boundaries bought recall no precision; long-event splitting duplicated an existing splitter and found no hypopnoea candidates; the coherence gate is a real repair of a real defect but costs 2.95/h of AHI bias on MESA when switched on. They ship as switches so the finding stays reproducible and the next cohort can overturn it. docs/third_party_comparison.md records one row per idea evaluated, including the ones rejected.

Architecture

~14,600 lines across 21 submodules, 817 unit tests (CI: Python 3.9–3.12):

constants · utils · signal · breath · breath_scoring · classify · spo2 · plm · ancillary · arousal · respiratory · indices · ventilation · pipeline · ml_classifier · profiles · postprocess · signal_quality · signal_quality_channels · ecg_effort · _types

Behaviour is pinned by a golden-output harness (PSGSCORING_GOLDEN=1) that scores fixed synthetic recordings and compares a digest. Every release runs it before publishing; a difference is a stop, not a re-bless.

Related

• YASAFlaskified — web platform integrating psgscoring with YASA staging, multilingual PDF reports, EDF+ export, and FHIR R4
• YASA — AI-based sleep staging (Vallat & Walker, eLife 2021)
• slaapkliniek.be — live instance (no installation required)

Citation

@article{rombaut2026psgscoring,
  title     = {Graded evidence in place of thresholds: an open-source,
               {AASM}-compliant method for respiratory event detection in
               polysomnography},
  author    = {Rombaut, Bart and Rombaut, Briek and Rombaut, Cedric},
  year      = {2026},
  note      = {Manuscript in preparation}
}

Disclaimer

psgscoring is research software — not a medical device. It is not CE-marked (MDR 2017/745) or FDA-cleared. All outputs are research-grade estimates that must be reviewed by a qualified clinician before any diagnostic or therapeutic decision. See DISCLAIMER.md for the full text.

License

BSD-3-Clause. See LICENSE.

Contact: bart.rombaut@gmail.com