Abstaining When the Clocks Disagree

When a prediction combines signals from two devices that keep their own clocks, such as a sleep-study recorder and a bedside monitor, the streams must first be lined up in time. The usual way is cross-correlation: slide one signal past the other and keep the time shift where they match best. That single shift is then treated as fact, and everything else about the alignment is thrown away. If noise, motion, or clock drift made the shift wrong, the downstream classifier receives inputs that are out of step with each other. These inputs can still look perfectly ordinary, so the classifier may answer with full confidence. This is a silent failure.

Selective prediction is the practice of letting a model say "I don't know" so that its error on the cases it does answer stays under a target. It normally takes its confidence from the classifier's own outputs, which is exactly where this failure is invisible.

We propose to measure how sure the alignment itself is and to use that as a second, independent reason to abstain. The measure comes from repeating the alignment over several pairs of channels shared by the two devices and checking whether their time shifts agree. This is where redundant sensor channels matter. Highly correlated channels agree with each other for the wrong reason, because they share the same noise. So the agreement only means something if it is computed over a deliberately non-redundant subset of channel pairs, weighted by how independent their errors really are. In two sentences: estimate the time offset between devices from several decorrelated channel pairs, and treat their disagreement as alignment uncertainty. Abstain when that uncertainty is high, even if the classifier is confident, and calibrate the combined rule so the error on accepted cases is controlled.

Core Claim

Our hypothesis is that, for cross-device fusion tasks where timing matters, alignment uncertainty catches a class of errors that the classifier's own confidence misses. If so, combining the two will give strictly better risk at a given coverage than either signal alone. We further hypothesize that how much this signal is worth depends on proxy-pair redundancy. Lag agreement among correlated channel pairs should be overconfident. Agreement among pairs chosen for independent errors should track true alignment error.

The idea sits on a gap the literature leaves open from both sides. On the alignment side, Time Synchronization of Multimodal Physiological Signals through Alignment of Common Signal Types cross-correlates one lead pair (Lead I against V2) on INCART ECG. It filters windows with a dynamic signal quality index and reports that accuracy collapses below about 6 dB SNR. However, it never turns alignment quality into a decision about a downstream prediction. OSASUD aligns a polysomnography recorder and a bedside monitor by maximizing cross-correlation on heart rate and SpO2. It then classifies every one-second window with no reject option, discarding the correlation landscape. On the abstention side, Selective Classification for Deep Neural Networks builds the selection rule from the maximum softmax probability and calibrates it with a binomial bound. COMPOSER rejects sepsis windows that look unlike a curated trust set, and it handles asynchrony with time-since-measurement weighting rather than alignment. Both look only at what the model or the input distribution reveals. Neither looks at whether the timeline itself is trustworthy.

Research Program

Alignment ensemble. For each prediction window, we compute normalized cross-correlation curves for a set of $K$ proxy pairs shared by the two devices. Examples are ECG leads recorded on both sides, or heart rate and SpO2 trends. Each pair gives a lag estimate $\hat\tau_k$ and a peak sharpness. This reuses the bandpass-plus-normalization pipeline of Xiao et al.

Measuring redundancy among proxy pairs. On clean calibration recordings with known offsets, we estimate the correlation matrix $C$ of the lag errors across pairs. Lag errors are what matter here, not the raw signals. We then select a subset $S$ by greedily maximizing $\det(C_S^{-1})$-style diversity, so that the chosen pairs have nearly independent errors. This borrows the determinant-based repulsion of Priors for Diversity in Generative Latent Variable Models. There it kept multiple latent components from redundantly modeling the same background; here it keeps several channel pairs from redundantly modeling the same noise.

The choice of errors over variance is deliberate. It follows Gaussian Image Anomaly Detection with Greedy Eigencomponent Selection, which showed that high-variance components carry background rather than task-relevant structure. The analogous risk here is that high-energy channels dominate the correlation sum and align on shared artifact.

Alignment uncertainty score. With weights $w = C_S^{-1}\mathbf{1} / (\mathbf{1}^\top C_S^{-1}\mathbf{1})$ (the generalized-least-squares combination of correlated estimates), the pooled lag is $\bar\tau = \sum_k w_k \hat\tau_k$. The alignment uncertainty is

$$u(x) = \sqrt{\textstyle\sum_{k\in S} w_k(\hat\tau_k-\bar\tau)^2 \,/\, (\mathbf{1}^\top C_S^{-1}\mathbf{1})^{-1}},$$

which is weighted lag dispersion scaled by the variance the pooled estimate would have if the calibration error model held. Values near 1 mean the pairs agree as well as expected; large values mean they disagree beyond what their known correlation explains.

Selection rule and calibration. The selection function accepts a window when $u(x)\le a$ and the softmax response $s(x)\ge b$. The thresholds $(a,b)$ are chosen on held-out patients with the selection-with-guaranteed-risk bound of Geifman and El-Yaniv, searched over a small grid with a union-bound correction.

Exchangeability across patients is assumed. Within-patient clock drift is the very thing being modeled, so we calibrate and test on disjoint patients and report the guarantee's empirical violation rate rather than assuming it holds.

Baselines

The naive reference fuses the two devices after single-pair cross-correlation and predicts on every window. This is the OSASUD pipeline. Its accuracy at full coverage anchors all curves.

The strongest competitor on selective risk is softmax-response selection with guaranteed risk, from Selective Classification for Deep Neural Networks, applied to the same fused classifier. On its own benchmark it reports 1% error at 78% coverage on CIFAR-10. There is no reported result on Lacuna for selective classification on OSASUD or INCART fusion tasks, so we run it ourselves. The proposal beats it if, at matched coverage on held-out patients, combined selection lowers selective risk with a confidence interval excluding zero. We also require it to cut the rate of accepted windows whose true residual lag exceeds the task tolerance.

The strongest alignment-side competitor is the signal-quality-index gate of Xiao et al.. It rejects windows whose waveform template match is poor, and reports a mean sync delay of 0.13 s at SQI = 1 while keeping 43% of the data, against 0.29 s without gating over 30-second windows. We use its gate as an abstention rule, alone and combined with softmax response. It is the most direct threat, because noisy windows are also badly aligned windows. Beating it means lower risk at the same coverage, especially on windows that are clean but drifted.

The robustness competitor trains the classifier with random time-shift and drift augmentation at the magnitudes we inject, then applies softmax-response selection. We also run an input-space trust check in the style of COMPOSER, a conformal distance to clean training windows.

Fair comparison. All methods share the same backbone, training data, augmentation budget, calibration split, and threshold-search grid.

Two ablations isolate the key ingredients. The first replaces the decorrelated subset with all proxy pairs weighted equally; this is the redundancy ablation. The second replaces the ensemble with the peak sharpness of the single best pair; this is the single-pair ablation. An oracle that aligns with the true injected offset bounds what alignment-aware abstention could gain.

Why not just…?

Why not just trust softmax confidence, since a misaligned input should look strange to the classifier? Sometimes it will. A misaligned ECG–PPG pair, however, is still a plausible ECG and a plausible PPG, and whether the classifier notices depends on whether it learned cross-device timing at all. We test this directly. We stratify accepted errors by true residual lag and ask whether softmax response ranks the misaligned windows below the aligned ones. If it does, adding alignment uncertainty buys nothing and we will report that.

Why not just gate on signal quality as Xiao et al. do? Quality gating removes noisy windows. Clock drift and an ambiguous correlation peak, as in periodic heartbeats where adjacent beats give near-equal peaks, can occur on clean signals. The quality index cannot see either. Our injection protocol separates noise from drift factorially. The prediction is that the quality gate matches us on noisy windows and loses on clean-but-drifted ones. If it matches us everywhere, the idea reduces to quality gating.

Why not just make the classifier robust to misalignment with time-jitter augmentation? This is the real threat. If the task tolerates the offsets that occur, abstaining for alignment is wasted coverage. We therefore report the jitter-trained model's accuracy as a function of injected lag, to establish whether the task is timing-sensitive at all. The proposal only claims gains where the jitter-trained model's accuracy degrades with lag. For beat-level fusion on INCART, at offsets comparable to a QRS width, we expect it to; for slow apnea trends on OSASUD, we may find it does not. That would be an informative negative result.

Why not just cross-correlate all channels and average? This is the redundancy ablation. Correlated leads should agree with each other even when all are wrong, inflating apparent certainty. If equal-weight all-pairs agreement performs as well as the decorrelated subset, the redundancy component is unnecessary and we drop it.

Proposed Experiments

We use two datasets with controlled desynchronization, since ground-truth offsets are otherwise unavailable.

On INCART, following Xiao et al., we split leads into a simulated "watch" (limb leads) and "patch" (precordial leads). We inject constant offsets, linear drift, and additive noise at SNRs from 0 to 20 dB. The downstream task is beat-type classification from fused features of both devices.

On OSASUD, we split the 18 channels by source device. We re-desynchronize the already-aligned recordings with the same injection grid and classify apnea windows. Patients are held out for calibration and test.

As a stress test of proxy selection, we use the 64-channel EEG of the driver drowsiness benchmark, where Zhou et al. found 54 channels redundant enough to hurt accuracy (78.26% full versus 81.38% with 10 channels). We partition the channels into two virtual headsets with independent clocks. This tests whether error-decorrelated proxy subsets differ from importance-selected ones, and whether they yield better-calibrated $u(x)$.

We report risk–coverage curves and area under them, selective risk at fixed coverages, and the empirical violation rate of the target risk across patient resamplings. We also report the silent-failure rate, defined as accepted windows with residual lag beyond tolerance.

The decisive check is whether the combined rule beats both softmax-response selection and quality gating, specifically on clean-but-drifted windows. It must do so on a task where the jitter-trained model is demonstrably timing-sensitive. The second decisive check is whether the decorrelated subset gives a smaller gap between predicted and actual lag error than all-pairs agreement. Failing the first falsifies the core claim. Failing the second removes the redundancy component but leaves the alignment-abstention idea intact.