A heart-sound classifier trained at one hospital often fails at the next, and much of the reason is mundane. A different stethoscope, a different chest-piece and a different room each change the recording's frequency response, much as a cheap speaker colors music. Models that read the raw waveform directly, the sampled pressure signal rather than a spectrogram, are particularly exposed to this. Their first layer is a learned bank of digital filters tuned to the training device. In certified clinical software, the model's weights also cannot be changed at the new site.
We propose to leave the model alone and correct the sound instead. At each new site we would estimate a short digital filter, an equalizer with a few dozen coefficients, that undoes the site's channel coloring before the audio reaches the frozen classifier. The equalizer would be estimated from unlabeled recordings, using the frozen model's own predictions (pseudo-labels) to compare like with like: target-site normal hearts against source-site normal hearts. This keeps a difference in disease prevalence from being mistaken for a difference in microphone.
Core Claim
Domain shift between stethoscope sites is dominated by a low-dimensional, roughly linear channel effect. If so, three things follow. The right adaptation target is that channel, not the model's features. The channel has to be estimated from class-conditional statistics, because matching overall average spectra would partly erase real disease signal when prevalence differs between sites. And pseudo-label noise becomes much less dangerous when it can only bias a filter with roughly 64 parameters, rather than being memorized by a network. The claim is testable: on held-out sites, a pseudo-label-informed equalizer in front of a frozen raw-waveform classifier should recover most of the gap to a target-trained oracle. It should do so without the confirmation-bias collapse seen when high-capacity adapters or full self-training absorb wrong labels.
Where This Sits
Raw-waveform learning supplies the model class. SoundNet showed that first-layer convolutions on raw audio learn frequency-selective, wavelet-like filters. Raw waveform CLDNNs reached parity with log-mel features for speech recognition. SincNet constrained those filters to parameterized bandpass shapes. All three evaluate in-domain and say nothing about sensor transfer. The research file flags the domain brittleness of raw front-ends as an open problem, and that brittleness motivates this work.
In medical acoustics, recent strong systems use spectrograms. These include the PhysioNet 2022 winner, a Mel-spectrogram plus tabular CNN; the semi-supervised Mean Teacher PCG model, which uses unlabeled PhysioNet 2016-e recordings through consistency training; and Spherical Procrustes Alignment, which fixes overconfidence on out-of-distribution medical audio by moving to cosine classifiers. Spherical Procrustes Alignment reduced calibration error on a BEATs backbone from 28.51% to 4.44%. Its authors state that its prototype updates assume clean labels. We borrow its diagnosis that softmax confidence under shift is untrustworthy, but we avoid prototype estimation from noisy labels altogether.
The adaptation constraint comes from Adapt the Data, Not the Model. Its INPUTADAPTER keeps a certified ICU predictor frozen and trains an axial-attention encoder-decoder in input space. It needs target labels, and it has enough capacity to learn arbitrary input rewrites. Our equalizer differs on both points: it is unsupervised and physically constrained. That paper also found its best adapter moved target data further from the source in Wasserstein distance. We read that as consistent with our view that marginal-distribution matching is the wrong objective, and our class-conditional estimate makes the alternative explicit.
The pseudo-label literature supplies the safeguards. Sener et al.'s reject option raised MNIST→SVHN accuracy from 32.3% to 40.3% by leaving uncertain targets out of the loss. Ensemble quality filtering and region uncertainty weighting down-weight dubious labels. Tri-net and Self-Ensembling Attention Networks document how errors compound when a network trains on its own mistakes. All of these protect a high-capacity learner. Our bet is that shrinking the learner removes most of the problem, so that rejection only has to keep a low-dimensional estimate unbiased.
Research Program
The source model is a raw-waveform heart-sound classifier with a SincNet-style first layer. It is trained on labeled source sites and then frozen. For a new target site we would fit a 64-tap finite-impulse-response equalizer $h$ placed before the frozen model. The fit proceeds in two stages.
The first stage is a closed-form initialization. For each class $c$, let $P^S_c(f)$ be the source long-term average power spectrum at frequency $f$. Let $P^T_c(f)$ be the same quantity on target recordings that the frozen model pseudo-labels as $c$ and that survive rejection. The channel estimate is
$$\log \hat G(f)^2 = \sum_c w_c \left[\log P^T_c(f) - \log P^S_c(f)\right],$$
where $w_c$ weights classes by the number of accepted pseudo-labels. The equalizer's magnitude response is set to $1/\hat G(f)$, realized as a minimum-phase filter.
The noise analysis follows directly. Suppose a fraction $\epsilon$ of recordings labeled $c$ actually belong to class $c'$. A first-order expansion of the log of a mixture spectrum gives a bias of about $\epsilon\,[\log P_{c'}(f) - \log P_c(f)]$. The bias is therefore bounded by the spectral difference between classes, which we expect to be much smaller than the differences between stethoscopes over most of the band. It is also largest where murmurs add high-frequency energy. This predicts that the estimate should lean on the majority, easier class (normal or absent), and that rejection matters most in the murmur band.
The second stage is optional refinement. We would backpropagate a pseudo-label cross-entropy, restricted to accepted samples, through the frozen model into the 64 taps. This lets the equalizer correct phase, which magnitude statistics cannot identify. The two stages alternate for two or three rounds.
Rejection uses a test suited to the confidence problem that Spherical Procrustes Alignment identified. A recording's pseudo-label is accepted only if it stays the same across a small ensemble of randomly perturbed equalizers drawn around the current estimate. This tests stability under channel uncertainty instead of trusting softmax magnitude. Every part of the procedure leaves the certified weights untouched.
Baselines
The naive reference is the frozen source model applied to target audio with no adaptation. Every gain is measured against it.
The strongest existing semi-supervised method on these data is the Mean Teacher PCG model. It updates model weights, which our setting forbids, so it serves as an upper-bound competitor rather than a like-for-like one. Its reported result on CirCor 2022 is a murmur score of 0.8180 and AUC 0.9004, using 80% labeled 2022 data and the 2016-e recordings as unlabeled data. We will rerun it on our splits with the same unlabeled target recordings and the same tuning budget. We count it as beaten if a frozen model plus equalizer matches it within confidence intervals, since our method changes no weights. Exceeding it would be a bonus, not a requirement.
The strongest frozen-model competitor is an INPUTADAPTER-style encoder-decoder (Gotfrid et al.) trained on our pseudo-labels instead of true labels. It gets the same rejection rule and the same number of optimization steps. The equalizer beats it if it is better on held-out-site modified accuracy, or equal while degrading less as pseudo-label noise rises.
We will also run classic channel normalization. For raw waveforms this means matching the target's overall long-term average spectrum to the source, which is the class-blind version of our first stage. For a spectrogram twin model it means per-band log-mel mean normalization.
The key ablation compares class-conditional estimation with class-blind estimation, at matched filter length. A second ablation compares rejection with no rejection. A capacity sweep runs from a single gain, to 64 taps, to 512 taps, to the full encoder-decoder, and isolates whether low capacity is what buys robustness to pseudo-label noise. An oracle equalizer fitted with true target labels marks the ceiling.
For context, the PhysioNet 2016 challenge winner reached modified accuracy (MAcc) of 0.8602 (Clifford et al.) on the official hidden test. That split mixes sites and is not a held-out-site result. We found no reported result on Lacuna for leave-one-source-out PhysioNet 2016 classification, so we will establish those numbers ourselves.
Why not just…?
Why not just normalize the average spectrum per site, with no pseudo-labels at all? This is the most serious threat. If the abnormal-recording rate at the target site is close to the source rate, class-blind matching and class-conditional matching should coincide, and pseudo-labels would add nothing. The PhysioNet 2016 subsets differ widely in their abnormal fraction, which gives us a natural test. We will plot the advantage of class-conditional over class-blind equalization against the target-minus-source prevalence gap. If the advantage does not grow with that gap, the pseudo-label component is unnecessary and we will report that.
Why not just use a Mel-spectrogram model with log-mel mean normalization? It may well work: that normalization is already a crude per-band equalizer, and Lu et al. won PhysioNet 2022 with Mel features. We will apply the same equalizer to a spectrogram twin of the source model, trained on identical data. The question then splits in two. One part is whether equalization helps both front-ends. The other is whether raw-waveform models gain more, because they are more channel-brittle. If raw waveform plus equalization does not beat Mel plus equalization, we will report that the raw front-end's value does not survive transfer.
Why not just self-train the whole model with confident pseudo-labels? Setting regulation aside, high-confidence thresholds under shift select confidently wrong predictions (Spherical Procrustes Alignment). Removing rejection cost Sener et al. eight accuracy points (Learning Transferrable Representations). We will run unconstrained threshold self-training at τ=0.95 as an unlocked reference. The noise-injection experiment below shows directly whether its accuracy degrades faster than the equalizer's as label noise rises.
Proposed Experiments
The primary benchmark is leave-one-source-out adaptation across the PhysioNet 2016 subsets (Clifford et al.). These subsets come from different groups and stethoscopes, and all carry normal/abnormal labels. We train on the remaining subsets, adapt using the held-out subset's unlabeled audio, and evaluate on its labels. Metrics are MAcc, sensitivity, specificity and expected calibration error, with recordings marked "Unsure" excluded from adaptation statistics.
The secondary benchmark uses CirCor 2022 murmur detection with 2016 recordings as unlabeled target data, following the Mean Teacher protocol, so that our numbers can be compared with its 0.8180 murmur score.
The decisive check is a controlled channel study. We will filter CirCor recordings through a family of synthetic stethoscope responses that also distort phase, so the true channel is known. We then inject pseudo-label noise at chosen rates. This lets us measure channel-estimation error directly against the first-order bias predicted above. It also lets us compare adapter capacities at each noise level.
The idea is supported if three things hold together: the equalizer's channel error tracks the predicted $\epsilon$-scaled bias, its target accuracy holds up flat over noise rates where the encoder-decoder adapter collapses, and the class-conditional advantage grows with the prevalence gap on real sites. It is refuted if class-blind normalization matches it on real sites, or if a Mel model with mean normalization matches raw waveform plus equalization.
References
- SoundNet: Learning Sound Representations from Unlabeled Video
- Learning the speech front-end with raw waveform CLDNNs
- Speaker Recognition from raw waveform with SincNet
- A Lightweight Robust Approach for Automatic Heart Murmurs and Clinical Outcomes Classification from Phonocardiogram Recordings
- Phonocardiogram (PCG) Murmur Detection Based on the Mean Teacher Method
- Spherical Procrustes Alignment for Reliable Medical Audio Diagnosis
- Adapt the Data, Not the Model: Input-Space Adaptation for Frozen Time-Series Predictors
- Learning Transferrable Representations for Unsupervised Domain Adaptation
- Auto-Annotation Quality Prediction for Semi-Supervised Learning with Ensembles
- Combating Noise: Semi-supervised Learning by Region Uncertainty Quantification
- Tri-net for Semi-Supervised Deep Learning
- Self-Ensembling Attention Networks: Addressing Domain Shift for Semantic Segmentation
- Classification of Normal/Abnormal Heart Sound Recordings: the PhysioNet/Computing in Cardiology Challenge 2016