acophenotype

End-to-end acoustic phenotyping: one recording in, a profile and a polished HTML report out.

acophenotype is the umbrella of the suite. It composes the other three packages — voxmarkers (interpretable biomarkers), acophen-streams (deep streams), and acophen-fusion (scoring) — into a single AcousticProfile and a self-contained HTML report.

⚠️ Research use only. Not a medical device. Not for diagnosis.

What you get

recording.wav
   │  (optional denoise: MetricGAN+)
   ▼
biomarkers (voxmarkers) ─┐
ssl · vision · emotion   ├─► acophen-fusion ─► prediction + per-stream contribution
(acophen-streams)        ┘
   ▼
AcousticProfile  ─►  .to_html("report.html")   ·   .to_json(...)

It degrades gracefully: with only the core installed you still get a biomarker-only profile (no fused prediction) — no errors, and the report adapts.

Install

pip install acophenotype

That's it. This installs the entire suite (voxmarkers, acophen-fusion, acophen-streams and the deep backbones). The model weights, VLAD codebook and reference cohort are downloaded automatically on first use and cached, so there is nothing else to configure.

Two genuinely optional add-ons:

pip install "acophenotype[denoise]"   # MetricGAN+ speech enhancement
pip install "acophenotype[report]"    # embedded log-mel spectrogram in the report

Quick start

from acophenotype import AcousticProfile

profile = AcousticProfile.from_audio(
    "recording.wav",
    task="binary",              # or "multiclass" / "regression"
    denoise=False,
    streams="all",              # or ["ssl", "vision"] etc.
    demographics=None,          # or {"sex":.., "age":.., "educ":..} for with_demos
)

profile.prediction              # "AD" / "CN" / None (biomarker-only)
profile.probability             # 0.78
profile.per_stream              # {"biomarkers":0.41, "embeddings":0.24, ...}
profile.percentile("temp_syllable_rate")   # vs the Pitt reference cohort
profile.to_html("report.html")  # the report
profile.to_json("report.json")  # machine-readable profile

CLI:

acophenotype analyze recording.wav -o report.html --json report.json
acophenotype analyze recording.wav --task regression --spectrogram

The report

A single self-contained HTML file — inline CSS and hand-drawn SVG, no external assets, no JavaScript, no matplotlib required. It renders:

• a colour-coded verdict + a semicircular P(AD) gauge,
• per-stream contribution bars (how each stream voted),
• biomarker percentile bars against the CN interquartile band, with out-of-range features flagged (each labelled with its clinical direction from voxmarkers),
• emotion aggregates with z-scores vs the reference cohort,
• a quality strip, provenance table, pipeline notes, and a prominent disclaimer.

This replaces the old run_7a/7b/7c scripts (monolithic extraction → report.json → matplotlib-base64 HTML). The extraction now comes from the suite packages, and the charts are crisp inline SVG that scale to any width.

Embed a log-mel spectrogram with profile.to_html("report.html", spectrogram=True) (needs the [report] extra).

Feature space (percentiles & flags)

Percentiles and flags are only meaningful when the patient vector and the reference cohort are in the same feature space. The reference shipped with the thesis (biomarker_cn_stats.json) is in standardized space while voxmarkers emits raw values — compared directly, nearly every feature flags. There are three paths, and the package handles all of them:

1. Regenerate the reference in your extraction space (recommended).

acophenotype build-reference \
    --features streams/biomarkers.parquet \
    --labels metadata.parquet --label-column target --control-value 0 \
    --space raw -o biomarker_cn_stats.json

2. Keep a standardized reference, but attach the scaler so raw vectors are converted automatically at scoring time:

from acophenotype import build_reference, scaler_from_features, save_reference
scaler = scaler_from_features(raw_features)          # per-feature mean/std
stats  = build_reference(standardized_features, space="standardized", scaler=scaler)
save_reference(stats, "biomarker_cn_stats.json")

With a scaler present, ReferenceData.align() z-scores the incoming raw vector into the reference space before comparing — no further action needed.

3. Neither of the above — the package detects the mismatch (via the fraction of features falling far outside the reference's observed range), records it in profile.notes, and shows a warning banner in the report instead of reporting misleading percentiles.

Weights & reference data (downloaded automatically)

Model weights live outside the wheel and are fetched from the Hugging Face Hub the first time you need them, then cached in ~/.cache/huggingface. Nothing to set up.

Resolution order for each asset: explicit argument → environment variable → local directory → automatic download.

If an asset genuinely can't be obtained, the profile still builds — it omits the parts that need it and records why in profile.notes.

License

MIT for the code. Pretrained backbones and model weights carry their own licenses.