How to interpret ProtoSleepNet#

This is the page to read to understand what the explanations mean. Unlike post-hoc attribution on a black box, ProtoSleepNet’s explanation and its prediction are the same object: the model can only reach a stage through the prototypes it activates.

1. The prototype codebook (Claim A — faithful by construction)#

Every 30-second epoch is encoded into a 128-dimensional embedding and then matched to the nearest of \(M=12\) prototypes by squared-L2 distance:

\[ k^\star(x) \;=\; \arg\min_{k \in \{1,\dots,12\}} \; \bigl\lVert\, \mathrm{encode}(x) - p_k \,\bigr\rVert_2^2 . \]

The stage is decided from these prototype activations. There is no surrogate model standing between the explanation and the decision — inspecting which prototypes an epoch resembles is inspecting the decision. This is what “faithful by construction” means, and it is why the atlas (a 2-D projection of the embedding space with the 12 prototypes placed inside it) is a legitimate map of the model’s reasoning, not a decoration.

Why the atlas is trustworthy

The 2-D atlas is produced with PaCMAP, fit jointly on epoch embeddings and the codebook. We report a nearest-prototype agreement — the fraction of epochs whose nearest prototype in 2-D matches their true nearest prototype in the full 128-D space. It is high (≈0.83–0.85 in-domain), so distances you see on the map reflect the distances that actually drive staging.

2. Per-epoch evidence (Claim B — traceable decisions)#

For any epoch you can ask why this prototype?. We answer with Integrated Gradients on the matching objective \(-\lVert \mathrm{encode}(x)-p_{k^\star}\rVert^2\) (zero baseline), which highlights the time–frequency regions that pull the epoch toward its prototype. In the demo this appears as a heatmap over the epoch’s spectrogram, per channel (EEG / EOG / EMG), next to the epoch’s place in the night’s hypnogram. Each decision is therefore traceable: a stage, a prototype, the evidence for the match, and the temporal context.

3. Clinical microstructure (Claim C & D — monosemantic, AASM-coherent)#

Each prototype carries a card: a 4-sentence natural-language rule, a spectral envelope, EEG band-relevance, channel-relevance, its dominant stage, and two quality scores — label purity and a monosemanticity score. High values mean the prototype is a clean, single-concept pattern (Claim C).

The signatures follow textbook AASM physiology (Claim D):

  • N3 prototypes concentrate on delta (0.5–4 Hz) EEG.

  • N2 prototypes concentrate on sigma / spindles (11–16 Hz).

  • REM prototypes are EOG-driven (eye movements) with theta.

  • Wake prototypes show alpha/beta EEG and EMG muscle tone.

4. The plausibility badge — an honest audit#

The demo shows a per-epoch plausibility badge: does the Integrated-Gradients relevance actually land on the frequency band and channel that the matched prototype’s stage predicts? This is a genuine sanity check of the local explanations, not a cherry-pick — epochs where the evidence lands off-band are shown as such, not hidden.

The audit below aggregates it across the featured recordings (anonymized “Recording A–D”, one set per model). “band ok” / “channel ok” are the fractions of epochs whose IG concentrates where physiology expects; “N3 pos” / “REM pos” are the mean normalized night-positions of N3 and REM prototypes (deep sleep should precede REM — and it does).

protosleepnet-gagliardi (seq)#

recording

epochs

plausible

band ok

channel ok

N3 pos

REM pos

Recording A

963

68%

81%

78%

0.36

0.83

Recording B

982

66%

83%

73%

0.32

0.62

Recording C

1011

65%

73%

74%

0.30

0.58

Recording D

940

68%

74%

81%

0.27

0.65

  • Recording A: N3 precedes REM (expected) — N3 mean position 0.36, REM 0.83.

  • Recording B: N3 precedes REM (expected) — N3 mean position 0.32, REM 0.62.

  • Recording C: N3 precedes REM (expected) — N3 mean position 0.30, REM 0.58.

  • Recording D: N3 precedes REM (expected) — N3 mean position 0.27, REM 0.65.

protosleeptransformer-gagliardi (st)#

recording

epochs

plausible

band ok

channel ok

N3 pos

REM pos

Recording A

848

77%

77%

100%

0.28

0.59

Recording B

990

63%

63%

100%

0.44

0.59

Recording C

1053

73%

73%

99%

0.36

0.52

Recording D

963

70%

70%

100%

0.27

0.57

  • Recording A: N3 precedes REM (expected) — N3 mean position 0.28, REM 0.59.

  • Recording B: N3 precedes REM (expected) — N3 mean position 0.44, REM 0.59.

  • Recording C: N3 precedes REM (expected) — N3 mean position 0.36, REM 0.52.

  • Recording D: N3 precedes REM (expected) — N3 mean position 0.27, REM 0.57.

Reading the numbers honestly

Per-epoch plausibility sits around 65–77%, not 100%. That is expected and intentional to report: Integrated Gradients marks relevance magnitude, not sign or a single mechanism, and real nights contain ambiguous, transitional epochs. What matters for the claims is that the evidence concentrates in the physiologically correct band/channel far more often than chance, and that the night-level structure (N3 before REM) is coherent for every recording.

5. The prototype-gram (Claim G)#

Rendering a whole night as its sequence of activated prototypes — colored by prototype rather than by stage — gives the prototype-gram, an intermediate representation that sits between the raw PSG and the coarse 5-class hypnogram. It exposes sub-stage structure (which kind of N2, which kind of REM) that the hypnogram flattens away.

Next: see it all together on the Live demo page, or regenerate the underlying analyses from Reproducing the paper.