A single neural network can be trained so that smaller pieces of it also work as standalone models. Such a network, often called a weight-sharing model bank, lets one training run serve a smartwatch, a phone, and a hospital workstation, each running a slice sized to its hardware. Medical signal models deployed this way face a second kind of variability that hardware-adaptation research ignores. Real electrocardiograms (ECGs, the electrical recordings of the heart) arrive with leads missing, electrodes misplaced, and motion noise, and the clinic needs to know when the model's answer should not be trusted. Two questions follow. Which slice of the bank should run on a given device and a given degraded recording? And how can that slice flag its own unreliable outputs without the 100–200 extra forward passes that current medical uncertainty methods require?
This note proposes an ECG model bank that can shrink in width, depth, and number of input leads. It also proposes a way to judge the bank's outputs that is free at inference time. The uncertainty signal is the disagreement between the early-exit prediction heads that the chosen slice already computes on its way to the final answer. This signal replaces the current label-free assumption that the largest model is the best judge, which we expect to fail under sensor corruption. The core claim is testable and specific in two parts. First, under realistic ECG corruption the largest sub-model stops being a reliable reference. Second, disagreement between exits in a single pass beats standard softmax entropy at rejecting wrong answers, at no more than about 10% extra latency on real mobile hardware.
What exists and where the gap is
The hardware side supplies the machinery. Slimmable Neural Networks trains one model that runs at several channel widths, with private normalization parameters per width, and shows that small widths beat identically sized standalone models by 3.3% on ImageNet. Layer Pruning on Demand with Intermediate CTC attaches auxiliary prediction heads to intermediate layers of a speech model so depth can be cut from 24 to 12 layers with no loss. One Proxy Device Is Enough for Hardware-Aware Neural Architecture Search shows that latency rankings transfer across devices after 30–80 target measurements. We reuse all three techniques: width switching, intermediate heads, and cheap latency transfer. All three treat the input as clean and fixed.
The closest cross-direction work is Slimmable Domain Adaptation (SlimDA). It trains a slimmable bank under domain shift and ranks sub-models on unlabeled target data by their output distance to the largest model, a label-free metric the paper calls UPEM. SlimDA reports that a 1/64-FLOP sub-model keeps 86.6% accuracy on ImageCLEF-DA versus 88.9% for the full model. Its ranking rests on an explicit "anchor assumption" that the biggest model is most accurate on the target. It was only tested on image-domain benchmarks, and it slims width uniformly rather than handling missing or corrupted input channels. We differ in three ways. We test the anchor assumption directly on corrupted physiological signals. We make the input lead set part of the elastic configuration. We replace distance-to-anchor with within-pass exit agreement for both selection and rejection.
Confident or Seek Stronger routes language-model queries between an edge model and the cloud using uncertainty. It finds that confidence distributions are shaped by the model more than by the data, and it explicitly leaves the effect of compression on calibration unstudied. That finding motivates our concern that every width/depth/lead configuration will have its own confidence profile. We borrow its idea of setting thresholds from held-out source data.
The robustness side supplies the input model and the evidence for the gap. Learning Cardiac Latent Representations in Vectorcardiogram Space (LVCG) projects multi-lead ECG into a physically grounded 3D heart-dipole space. It trains by reconstructing masked leads, and it stays stable when input leads drop from 12 to 3. That makes it our backbone, because a lead-count switch becomes natural in dipole space. Its single-dipole physics is also a known limitation that our corruption tests will probe. Melanoma Detection with Uncertainty Quantification shows that rejecting the 20% most uncertain cases by softmax entropy raises accuracy from 93.2% to 97.8%. Uncertainty Estimation in Deep Neural Networks for Dermoscopic Image Classification reaches 0.80 out-of-distribution detection AUC, but only by using 100 dropout passes plus 100 augmented passes. That cost is exactly what edge hardware cannot pay; OODIn documents that phones throttle and migrate workloads within about a second. Practical Strategies for Extreme Missing Data Imputation in Dementia Diagnosis constrains the design: models that use only the features present beat imputation (AUC 0.896). We therefore handle missing leads by masking rather than filling them in.
Core Claim
We make three hypotheses. First, on corrupted ECGs (lead loss, electrode swaps, baseline wander, motion noise), the largest configuration of a slimmable bank is frequently not the most accurate. As a result, SlimDA-style anchor ranking correlates poorly with true sub-model accuracy. Second, the agreement between a configuration's intermediate exit heads and its final head is a better signal for rejecting wrong answers than calibrated entropy, specifically under corruption. The reason is that corruption tends to make shallow and deep features disagree even when the final softmax stays confident, and this signal is available in one pass. Third, the same agreement signal ranks configurations without labels more faithfully than distance-to-largest-model.
Research Program
We build the model bank on an LVCG-style encoder. Each available lead is embedded and projected toward the 3D dipole space, cardiac cycles are compressed into tokens, and a recurrent layer models rhythm. We make this encoder elastic along three axes: channel width (four switches, with private normalization per switch), depth (exit heads after each quarter of the layers, trained jointly in the style of intermediate-CTC supervision), and input leads (12, 6 limb, 3, 1). Training samples a random (width, depth, lead-subset) configuration per step, alongside the largest and smallest ones, and keeps LVCG's masked-lead reconstruction loss. We pretrain on MIMIC-IV-ECG and fine-tune on PTB-XL.
For a configuration with exits $1..K$, the uncertainty score is
$$u(x) = H(p_K(x)) + \lambda \cdot \frac{1}{K-1}\sum_{k<K} \mathrm{JS}\big(p_k(x),\, p_K(x)\big).$$
Here $p_k$ is the predicted label distribution at exit $k$, $H$ is entropy, $\mathrm{JS}$ is Jensen–Shannon divergence, and $\lambda$ is a weight fixed on source validation data. The rejection threshold for each configuration is set on held-out PTB-XL to hit a target coverage, then frozen before any shifted data is seen.
For label-free selection on a target site, each configuration is scored by the mean of $u$ over unlabeled target recordings. The score is traded off against latency predicted by an AdaProxy-style transfer from a proxy phone.
Baselines
The naive reference is a single fixed-size LVCG model, one per lead count, with plain softmax entropy rejection and one threshold. This is what a practitioner following Melanoma Detection with Uncertainty Quantification would ship.
The strongest cheap competitor is our own bank with per-configuration temperature scaling followed by entropy rejection. This is the obvious fix for calibration drift, so it isolates whether exit disagreement adds anything beyond recalibration. It uses the same checkpoints, the same held-out calibration split, and the same coverage targets.
The strongest existing label-free selection method is SlimDA's UPEM applied to our bank. It is the only published approach to ranking sub-models without labels, and it is run on identical target data.
The strongest existing uncertainty method is MC Dropout plus test-time augmentation following Combalia et al.. We run it at its reported 100+100 passes as an accuracy ceiling. We also run it at a pass count matched to our measured latency budget (likely 1–2 passes), which is the fair edge comparison. A deep ensemble of independently trained small models at matched parameter count completes the set.
For reported numbers, the relevant ones are the following. The reference OOD AUC is 0.80 on ISIC 2018 (Combalia et al.). LVCG reports 72.03% AUC on PTB-XL Rhythm with 1% labels and a margin of more than 10% AUC over HeartLang on CPSC 2018. No reported result found on Lacuna for selective prediction, calibration, or UPEM ranking on ECG slimmable models; our baselines establish those numbers. Beating a baseline means a lower area under the risk-coverage curve on corrupted CPSC 2018, with bootstrap confidence intervals excluding zero, at no more than 1.1× single-pass latency on device.
The key ablations remove the exit-disagreement term (λ=0), remove lead-masked training, and replace the dipole projection with a plain 12-channel convolutional encoder. The last one tests whether the physical bottleneck is what keeps smaller configurations robust.
Why not just…?
Why not just temperature-scale each configuration and use entropy? This is the real threat. If corruption mostly shifts confidence uniformly, recalibration will match our method. The per-configuration temperature-scaling baseline settles it directly. If it ties our score under corruption, we will report that the drift is fixable by recalibration and drop the disagreement term. The anchor-failure and selection results do not depend on the disagreement term and would still stand.
Why not just trust the largest model, as SlimDA does? SlimDA never tested physiological corruption. We measure, per corruption type, how often the largest configuration loses to a smaller one and the Spearman correlation between UPEM and true accuracy. If the anchor holds, UPEM wins selection and we say so.
Why not just train separate models per lead subset, as the reduced-feature dementia work suggests? McCombe et al. show per-subset models beat imputation, and we agree about not imputing. But subsets multiplied by widths yields dozens of models, and weight sharing beat standalone small models in Slimmable Neural Networks. We include independently trained per-lead-subset models at matched size as a baseline, so this claim is tested on ECG rather than assumed.
Why not just use a few MC Dropout passes? At matched latency, the budget allows roughly one extra pass, which yields an extremely noisy variance estimate. The latency-matched MC baseline measures this directly.
Proposed Experiments
We evaluate on PTB-XL (source), and on CPSC 2018 and Chapman-Shaoxing as shifted sites, all from the LVCG benchmark suite. Controlled corruptions are applied at test time: lead removal down to 3 and 1 leads, limb-electrode swaps and sign flips, baseline wander, and synthetic motion noise. The motion noise is simulated because the file lists no real ambulatory-noise ECG corpus, and we flag this as a limitation. The metrics are macro AUROC, expected calibration error per configuration, area under the risk-coverage curve, missed-positive rate among accepted cases at 80% coverage, and Spearman correlation between label-free and true configuration rankings. Latency is measured on a mid-range phone CPU and NPU and on an embedded board, with AdaProxy-style transfer checked against direct measurement.
Three checks decide the project. The first is anchor failure: whether the largest configuration loses to a smaller one on a substantial fraction of corrupted conditions, and whether UPEM's rank correlation drops well below its clean value. The second is calibration drift: whether a single frozen threshold yields coverage that swings across configurations and lead counts, and whether per-configuration thresholds fix this. The third is the headline comparison: whether exit disagreement beats per-configuration temperature-scaled entropy on corrupted CPSC 2018 at matched latency. If the first check fails, we still contribute the first measurement of calibration and selection for elastic ECG models. If the third fails, the simple recalibration baseline is the right answer and we report it as such.
References
- Slimmable Neural Networks
- Layer Pruning on Demand with Intermediate CTC
- One Proxy Device Is Enough for Hardware-Aware Neural Architecture Search
- Slimmable Domain Adaptation
- Confident or Seek Stronger: Exploring Uncertainty-Based Small LM Routing From Benchmarking to Generalization
- LVCG: Learning ECG Representations in the Latent Vectorcardiogram Space
- Melanoma Detection with Uncertainty Quantification
- Uncertainty Estimation in Deep Neural Networks for Dermoscopic Image Classification
- OODIn: An Optimised On-Device Inference Framework for Heterogeneous Mobile Devices
- Practical Strategies for Extreme Missing Data Imputation in Dementia Diagnosis