A nurse with a digital stethoscope in a crowded rural clinic, with no cardiologist anywhere nearby, needs three answers from an automated assistant: "normal, no referral needed," "refer," or "this recording is too noisy, record again." The third answer is the one current systems handle worst. Murmur classifiers that win public challenges still struggle to flag degraded recordings. They also fail sharply once background noise approaches the loudness of the heart sounds. The dangerous result is a confident "normal" on a recording that contains almost no cardiac information. We propose to borrow a tool from radio engineering to make the "record again" decision reliable. Cognitive radios must decide whether a weak transmitter is present when the noise level itself is uncertain. That field showed that any test based on loudness alone hits a hard floor, the so-called SNR wall (signal-to-noise ratio wall). Tests that look for the signal's periodic structure avoid that floor.
The core idea is easy to state. A heartbeat is periodic, and most clinic noise is not periodic at cardiac rates. We will therefore measure how much beat-locked periodic structure a recording contains, using a statistic that does not change when the noise gets louder or quieter. We will calibrate its threshold against a null distribution, meaning the distribution of the statistic when only noise is present. A recording passes to the murmur classifier only when the statistic is high enough, and the threshold is set so that noise-only recordings are wrongly accepted at a fixed, known rate. This is the medical analog of a constant-false-alarm-rate (CFAR) detector. The hypothesis is that this structure-based gate keeps the miss rate among accepted recordings bounded across noise types and levels where confidence-based abstention does not.
Core Claim
Gating triage decisions on a scale-invariant measure of cardiac periodic structure, with a calibrated null distribution, will control the false-normal rate on accepted heart recordings under unknown and shifting noise. It will do this more reliably than thresholding the classifier's own confidence or using a learned quality classifier. The control should hold especially for noise types absent from training.
The claim rests on a result from spectrum sensing. Spectrum Sensing in Cognitive Radios Based on Multiple Cyclic Frequencies builds a generalized likelihood ratio test (a test comparing how well "signal present" and "noise only" explain the data). It works by summing evidence of cyclic correlation across several cyclic frequencies, meaning repeating correlation at a waveform's built-in periodicities. That paper reports reliable detection down to −15 dB with cooperation, and its sum statistic has a chi-square null distribution, which allows analytic thresholds. We reuse this method. What changes is the application: the periodicity is the cardiac cycle and its harmonics rather than a carrier or symbol rate, and the cycle rate is unknown and drifts.
Two further spectrum-sensing papers explain why energy-based gates are the wrong design. Deep Learning for Spectrum Sensing shows energy detection failing near −6 dB while a structure-learning network reaches about −8.5 dB. Enhanced Detection Algorithms Based on Eigenvalues and Energy in Random Matrix Theory Paradigm shows energy detectors collapsing under only 2 dB of noise uncertainty. Gating heart sounds on estimated loudness or SNR inherits this wall.
On the clinical side, three papers document the gap. Exploring the Impact of Noise and Degradations on Heart Sound Classification Models mixes 16 clean recordings with 21 noise types at SNRs from −10 to +40 dB. Its classifiers degrade below +10 dB and collapse below 0 dB, with ambient and colored noise the most damaging. A Lightweight Robust Approach for Automatic Heart Murmurs and Clinical Outcomes Classification from Phonocardiogram Recordings wins the 2022 murmur challenge with a segmentation-free spectrogram network and heavy augmentation, yet it struggles on the "Unknown" class. Heart Sound Classification Using Deep Structured Features likewise degrades when the "Unsure" label is added. None of these sets a noise-invariant, calibrated threshold for abstaining.
Relation to the Closest Work
The nearest existing work treats noise as something to measure or train against, not as an unknown nuisance with a guaranteed test. Assessing Robustness to Noise: Low-Cost Head CT Triage simulates low-cost scanner physics and retrains a triage network on the degraded scans. It reports robustness, but it gives no rule for when to refuse a decision. Shariat Panah et al. characterize where accuracy fails but propose no detector. Lu et al. rely on augmentation.
The design closest to ours in spirit is the uncertainty-gated ultrasound system in Fetal gestational age estimation using artificial intelligence on non-targeted ultrasound images and video. That system lets novice operators sweep freely and uses a Kalman filter to reject frames whose learned uncertainty is too high, reaching a confident estimate in a median of 24 seconds. We adopt its workflow logic of rejecting and re-acquiring until confident. We replace its learned uncertainty head with a statistic whose noise-only behavior is known by construction. The learned head has no guarantee on noise it never saw, and that is exactly the failure we target.
The setting is also deliberately narrower than Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs. That study sets triage operating points empirically on a million chest X-rays and reports an 80% workload reduction at a 1.57% false-negative rate. Its operating points come from validation curves in one low-prevalence center. Ours come from a null model that does not depend on the noise level.
We considered and set aside the other obvious bridge, eigenvalue-ratio detection across the four auscultation sites in the style of Cooperative Spectrum Sensing Using Random Matrix Theory. That detector needs simultaneous samples from several sensors. The heart-sound datasets in the file describe separate recordings per valve location, not a synchronized array, so the cross-site covariance has no physical meaning.
Research Program
The statistic comes first. Each recording is band-pass filtered into a low band, where S1 and S2 dominate, and a higher band, where murmurs live. We then take the amplitude envelope $e_b(t)$ of each band $b$. For a candidate heart rate $\alpha_0$ in the physiological range (roughly 0.7 to 3.5 Hz, widened for pediatric rates), we estimate the cyclic autocorrelation of the envelope at harmonics $k\alpha_0$, $k = 1,\dots,K$, and a set of lags. Following Lundén et al., we stack these estimates into a vector $\hat{r}_b(\alpha_0)$, estimate its covariance $\hat{\Sigma}_b(\alpha_0)$ from the same recording, and form the sum statistic
$$T_b(\alpha_0) = \hat{r}_b(\alpha_0)^{H}\,\hat{\Sigma}_b(\alpha_0)^{-1}\,\hat{r}_b(\alpha_0).$$
Here $H$ is the conjugate transpose. Because $T_b$ is a quadratic form normalized by its own estimated covariance, multiplying the noise by any constant leaves it unchanged. This is the property that defeats the SNR wall for a scalar noise level. The recording-level evidence is $T^\ast_b = \max_{\alpha_0} T_b(\alpha_0)$.
The maximum over an unknown, drifting heart rate breaks the chi-square result. We therefore calibrate the null with surrogates: we randomize the Fourier phases of $e_b(t)$, which keeps its power spectrum but destroys cyclic structure, and recompute $T^\ast_b$. Tolerance to slow heart-rate drift comes from short overlapping windows with a smoothness penalty on $\alpha_0$ across windows. The low-band statistic answers whether enough heartbeat is present to judge. The high-band statistic, evaluated at the $\alpha_0$ found in the low band, measures beat-locked high-frequency energy, which is a noise-scale-invariant measure of murmur evidence.
The triage system has three outputs. A recording whose low-band evidence falls below the surrogate threshold, set for a target false-acceptance rate $\epsilon$ on noise-only input, is returned as "record again." Accepted recordings go to a murmur classifier. That classifier is a reimplementation of the Lu et al. segmentation-free network, given the cyclic statistics as extra inputs, and it outputs "refer" or "normal." Its normal/refer threshold is chosen on accepted calibration recordings to hit a target miss rate.
The analysis question is whether that miss rate stays stable when noise type and level shift at test time. For the noise-only false-acceptance rate, the surrogate construction gives a guarantee. For the miss rate among accepted recordings, we can only test it empirically.
Baselines
The naive reference point is the Lu et al. classifier with no abstention at all, run on every recording. On the 2022 challenge hidden test set Lu et al. report a murmur score of 0.780, ranked first, according to A Lightweight Robust Approach for Automatic Heart Murmurs and Clinical Outcomes Classification from Phonocardiogram Recordings. Because the hidden set is not available to us, we will retrain it with patient-level cross-validation on the public CirCor DigiScope recordings and report our own reproduction alongside that figure rather than comparing against it directly. On PhysioNet 2016 the corresponding reference is the winning entry's modified accuracy of 0.8602 with sensitivity 0.9424, from Classification of Normal/Abnormal Heart Sound Recordings: the PhysioNet/Computing in Cardiology Challenge 2016, again on a hidden set. The file contains no noise-stratified or risk–coverage numbers for any of these systems (no reported result found on Lacuna for risk–coverage of heart-sound murmur classifiers), so those curves will be produced here for every method.
The strongest competitor for the gate is selective prediction on the same classifier. Its confidence is thresholded, and the threshold is calibrated by split-conformal risk control to the same target miss rate on the same calibration recordings. This is strong because it needs no new signal processing, it adapts to whatever the classifier learned from augmentation, and on in-distribution noise it may match us exactly. The second strong competitor is a learned quality classifier trained on the "Unknown" and "Unsure" annotations, which is what a practitioner would build. The third reference is the radio-world strawman: a gate on estimated SNR or band energy. We expect it to hit a wall, and including it tests whether that transfer of intuition is correct.
Fairness is enforced by sharing everything except the gate. All methods use the same backbone, the same augmentation recipe, the same training folds, the same calibration split, and the same hyperparameter search budget, with gate thresholds tuned to the same target $\epsilon$ and miss rate. Beating the strongest baseline means one of two outcomes at matched coverage on held-out noise types. Either the miss rate among accepted recordings stays within its target while conformal confidence gating exceeds it, or our coverage is higher at the same realized miss rate.
Three ablations isolate the key ingredient. The first replaces the covariance-normalized statistic with raw cyclic magnitude, which removes scale invariance. The second replaces surrogate calibration with the asymptotic chi-square threshold, testing whether the max over heart rate matters. The third feeds the cyclic features to the classifier but removes the gate, separating "better features" from "better abstention."
Why not just…?
Why not just augment more and trust the classifier, as Lu et al. did? Augmentation teaches robustness to the corruptions that were simulated. The evidence suggests it does not move the operating floor. Shariat Panah et al. find that even their best, transfer-learned spectrogram network collapses below 0 dB. Lu et al.'s own winning system is weakest on the "Unknown" recordings, the cases that matter for a safety gate. Our held-out-noise-type split tests this directly: we augment with some noise families and test on others.
Why not just threshold the classifier's confidence with conformal calibration? This is the real threat. Conformal guarantees hold only when calibration and test data are exchangeable. A clinic whose noise differs from the calibration set breaks that assumption, while our noise-only false-acceptance bound holds for any stationary noise of any level. Whether that difference matters in practice is an empirical question, and the held-out-noise and SNR-sweep experiments will settle it. If confidence gating matches the cyclic gate even on unseen noise types, then the gate adds nothing. The contribution would shrink to the cyclic features, and we would report it that way.
Why not just train a quality classifier on the "Unknown" labels? Those labels reflect annotator judgment rather than information content, and they are scarce. Both Lu et al. and Tschannen et al. report poor performance on the noisy category. A learned quality model is also exposed to the same shift problem as confidence gating. We will measure agreement between each gate and the annotator labels, and compare their miss-rate control under shift.
Why not just denoise first, or gate on estimated SNR? Shariat Panah et al. find that improving SNR beyond about +20 dB gives no accuracy benefit. Filters that share a band with murmurs risk deleting them. Gating on loudness is, in the radio literature, precisely the design that fails under noise uncertainty (Gao et al.; Zhao et al.). The SNR-gate baseline should show this. If it does not, the transfer from radio is weaker than claimed.
Proposed Experiments
The main benchmark is the CirCor DigiScope recordings from the 2022 challenge, about 5,282 pediatric recordings from 1,568 patients with Present, Absent and Unknown murmur labels. PhysioNet 2016 serves as a second, adult, multi-source set with Normal, Abnormal and Unsure labels.
On both, we add noise at controlled SNRs from −10 to +40 dB using the Shariat Panah et al. noise taxonomy of colored, movement, internal-physiological and ambient noise, applied to the full datasets rather than 16 recordings. Training sees only some noise families; testing uses the rest. We also introduce deliberate noise-level drift within a recording, to mimic an uncalibrated device. Reported metrics are the miss rate on murmur-present cases among accepted recordings, coverage, the false-acceptance rate on pure-noise and heartbeat-free recordings, the re-record rate, and weighted murmur accuracy on accepted cases, all plotted against SNR and noise family.
Three checks decide the outcome. First, on noise-only recordings at every noise level, the cyclic gate's false-acceptance rate should stay near the target $\epsilon$. The confidence and SNR gates should drift with noise level. If ours drifts too, the invariance claim fails. Second, on held-out noise families, the miss rate among accepted recordings under our gate should stay within target where conformal confidence gating exceeds it. A tie here means we drop the gate claim. Third, on the original, clean-ish recordings, our gate should reject few recordings the annotators considered usable, so that the safety benefit does not come from simply refusing most inputs. A final stress test injects rhythmic non-cardiac noise, such as a crying infant or periodic machinery, near cardiac rates. This is the known blind spot of any periodicity-based test, and we will report how often it fools the gate.
References
- Spectrum Sensing in Cognitive Radios Based on Multiple Cyclic Frequencies
- Deep Learning for Spectrum Sensing
- Enhanced Detection Algorithms Based on Eigenvalues and Energy in Random Matrix Theory Paradigm
- Exploring the impact of noise and degradations on heart sound classification models
- A Lightweight Robust Approach for Automatic Heart Murmurs and Clinical Outcomes Classification from Phonocardiogram Recordings
- Heart Sound Classification Using Deep Structured Features
- Assessing Robustness to Noise: Low-Cost Head CT Triage
- Fetal gestational age estimation using artificial intelligence on non-targeted ultrasound images and video
- Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs
- Cooperative spectrum sensing using random matrix theory
- Classification of Normal/Abnormal Heart Sound Recordings: the PhysioNet/Computing in Cardiology Challenge 2016