Predicting Which Wearable Heart-Rate Networks Survive Motion, Before Training Them

Wrist-worn heart-rate sensors work well while you sit still and badly while you run. The pulse appears as a tiny change in reflected light. Arm swing and footfalls distort that light far more, and they do so at 1–3 Hz, the same band as the heart rate itself. The networks that cope best are small one-dimensional temporal convolutional networks (TCNs: stacks of causal convolutions whose "dilation" spaces out the kernel taps so each layer sees further back in time). They have to fit in under a megabyte on a microcontroller, so choosing their width, depth, and dilation schedule is a real search problem. Today that search is done by training: tools that grow and prune channels or dilations during gradient descent. Meanwhile, the computer-vision community has built "zero-cost proxies" that score an untrained network from one or two gradient passes at random initialization and rank architectures nearly for free. Nobody has checked whether these proxies say anything about wearable time-series models, or about robustness to motion in particular.

This note proposes to answer that question with a purpose-built benchmark of fully trained PPG heart-rate TCNs. It also tests one specific new proxy. The proxy rests on a simple bet. A network is likely to learn motion-invariant features if, at initialization, the gradients it receives from motion-corrupted windows already point the same way as the gradients from resting windows with the same true heart rate. Existing proxies measure how strong and consistent gradients are across a random batch. Ours measures whether they agree across physical activity states. If the bet holds, architectures could be screened for motion robustness, not just clean accuracy, in seconds. If it fails, the benchmark still shows where training-free estimation breaks for physiological sensing.

Core Claim

We claim that rank correlation between a proxy and final heart-rate error depends on which windows the proxy is computed on and on which error we try to predict. On motion-heavy activities and at activity transitions, standard proxies computed on randomly sampled windows will rank architectures no better than a simple structural model built from parameter count and receptive-field length. A proxy defined as gradient agreement between matched rest and motion windows will do better on exactly those error targets. The claim is falsifiable. It rests on a benchmark that does not yet exist, so building that benchmark is half the contribution.

Where This Sits

Prior PPG work supplies the application and the training-based competitors. TimePPG (Robust and Energy-Efficient PPG-Based Heart-Rate Monitoring; Embedding Temporal Convolutional Networks for Energy-Efficient PPG-Based Heart Rate Monitoring) uses MorphNet, a pruning method that learns channel importance during training, to trace error-versus-cost Pareto fronts on PPG-DaLiA. Q-PPG adds Pruning-In-Time to learn dilations, followed by mixed-precision quantization. Both pay for every search decision with training, and both optimize clean aggregate loss rather than anything motion-specific. We keep their search space and their deployment constraints and replace the selection signal.

Zero-cost-proxy work supplies the method and also constrains it. The PPG proxies we reuse come from four papers:

  • Zero-Cost Proxies for Lightweight NAS contributes the pruning-at-initialization proxies (snip, synflow, grad_norm) and Jacobian covariance.
  • ZiCo scores the per-layer ratio of mean to standard deviation of absolute gradients.
  • AZ-NAS combines expressivity, trainability, and complexity proxies by rank aggregation.
  • TEG-NAS grounds expressivity and trainability proxies in linear-region counts and neural-tangent-kernel conditioning.

All four are validated on vision benchmarks. ZiCo itself reports that many proxies fail to beat raw parameter count. That result is the main constraint on our design, and it is why our strongest naive baseline is structural rather than random.

Two robustness papers are the closest existing work. An Evaluation of Zero-Cost Proxies shows that adversarial robustness on NAS-Bench-201 is much harder to predict from proxies than clean accuracy, with R² around 0.65 against roughly 0.97. Zero-cost Proxy for Adversarial Robustness Evaluation builds a proxy from the smallest tangent-kernel eigenvalue and input-loss curvature, and that proxy penalizes skip connections.

Both papers motivate us: robustness is a distinct prediction target. Both differ from us in kind. Their perturbations are small worst-case image perturbations. Ours are real, large, low-frequency motion artifacts that overlap the target band. The skip-connection bias is also a direct threat, because residual connections are standard in TCNs. We therefore treat the adversarial proxy as a baseline we expect to mislead, not as a template.

Finally, Change Point–Aware Evaluation and Ambulatory Atrial Fibrillation Monitoring show that dataset-wide error hides failures at activity transitions. That motivates scoring proxies against transition error, not only overall MAE.

Research Program

The first deliverable is a tabular benchmark, which we call PPG-TCN-Bench.

Search space. We define a TCN space around the TimePPG/Q-PPG seed. It varies depth (3–9 blocks), per-block width, kernel size, per-layer dilation, whether each block has a residual connection, and whether accelerometer channels enter early or late. We filter candidates to those that fit STM32L4/STM32WB55-class memory.

Training protocol. We sample about 1,000 architectures stratified by parameter count and receptive field. Each is trained to convergence on PPG-DaLiA, with three seeds and a fixed leave-subjects-out four-fold protocol. All candidates share one training recipe and the physiological post-processing filter used by TimePPG.

Recorded outcomes. For every trained architecture we record overall MAE, per-activity MAE, and MAE in the 30 seconds after each annotated activity change. We also record MACs and memory, plus measured latency and energy for a subset deployed on hardware. These models have roughly 5k–100k parameters and the dataset is 37.5 hours, so each run costs minutes. The full table is affordable and reusable.

The motion-agreement proxy. From PPG-DaLiA we draw matched window pairs. Each pair consists of a rest window and a motion window, taken from the same subject, whose ECG-reference heart rates differ by less than 3 BPM. Motion is defined by accelerometer energy. For a randomly initialized network with weights $\theta_0$ and layers $l$, let $g_r^l$ and $g_m^l$ be the layer-$l$ loss gradients averaged over the rest and motion halves of a batch. The score is

$$A(\theta_0) = \sum_l \frac{\langle g_r^l, g_m^l \rangle}{\lVert g_r^l \rVert \, \lVert g_m^l \rVert} \cdot \log \frac{\text{mean}|g^l|}{\text{std}|g^l|},$$

where the second factor is ZiCo's per-layer signal-to-noise term computed over the pooled batch.

The cosine term asks whether early learning is pushed the same way by an artifact-laden view of a heartbeat and a clean view of the same heartbeat. The signal-to-noise term guards against layers whose gradients agree only because they are vanishingly small. We report each factor alone as an ablation. We also average $A$ over several initialization seeds and report its seed variance, since a proxy that is noisier than the gaps it ranks is useless.

Transfer test. We take the top-ranked architectures under each proxy, retrain them, and test them, without further search, on the Tri-Spectral PPG wild-activity data and on SPC Cup 2015. This checks whether proxy-based selection overfits DaLiA's activity mix.

Baselines

Structural reference points. The do-nothing reference is random selection within the memory budget. The naive reference a reviewer will actually press is parameter count. We strengthen it into a structural regressor: a small model fit on parameter count, MACs, and analytic receptive-field length, trained on a held-out slice of the benchmark. Receptive field is free to compute. Q-PPG's dilation stage suggests it matters under motion, so this is the honest "no gradients needed" competitor. Any proxy must beat it in rank correlation on motion and transition error, with bootstrap confidence intervals over architectures and initialization seeds that do not overlap. A proxy whose gain disappears once receptive field is controlled for has not earned its keep.

Existing zero-cost proxies. We run ZiCo, snip, synflow, Jacobian covariance, the tangent-kernel condition number, the AZ-NAS rank ensemble, and the adversarial proxy $R$. All are computed with the same batch size, the same number of passes, and the same seeds as $A$. The strongest of these on vision is AZ-NAS, which reports Kendall's τ of 0.710 on NAS-Bench-201 ImageNet16-120. ZiCo reports τ of 0.61 on NATSBench-TSS against 0.57 for parameter count. We found no reported result on Lacuna for any zero-cost proxy on PPG or other one-dimensional physiological data, so the in-domain bar is set by our own benchmark.

Key-ingredient ablation. The most important ablation computes every standard proxy on motion-only batches and on mixed random batches. This separates "agreement across activity states" from "merely seeing corrupted data."

Training-based search. The strongest end-to-end competitor is the Q-PPG pipeline, which reports 4.36 BPM MAE on PPG-DaLiA. TimePPG reports 4.88 BPM, or 3.84 BPM with subject-specific fine-tuning (TimePPG). We rerun MorphNet+PIT on our folds with our training recipe. Proxy-guided search wins if its selected model is within 0.3 BPM of the MorphNet+PIT model at matched memory, for at least an order of magnitude less search compute. It wins outright if it is better on motion-window error.

Why not just…?

Why not just use parameter count, or parameter count plus receptive field? This is the most serious threat. ZiCo's own results show plain parameter count is hard to beat, and in a narrow TCN space a structural model may capture most of the variance. We do not argue around it. The structural regressor is the primary baseline, and the paper's claim is stated relative to it. If it matches $A$ on motion and transition error, we report that proxies add nothing for this domain. That would still be a useful, publishable negative result for wearable designers.

Why not just run MorphNet or Pruning-In-Time as TimePPG and Q-PPG do? Those methods work: 3.84–4.36 BPM is the bar. Their cost, though, is a full training run per search configuration, and they explore width and dilation around a single seed rather than macro-topology. Our search-cost-versus-MAE comparison at matched memory measures directly whether the proxy route reaches the same Pareto front. If it does not, the gap tells us how much training signal the proxies are missing.

Why not just train each candidate for one epoch and rank by early validation error? Short training is a strong and cheap ranking signal, and it naturally sees motion data. We include one-epoch and five-epoch early error as baselines with compute measured in matched GPU-seconds. We expect them to beat all proxies on overall MAE. The open question is whether early error, which is dominated by easy resting windows, predicts transition error as well as $A$ does. The benchmark answers this directly.

Why not just switch between models with a motion detector, as ActPPG does? ActPPG halves compute for a 0.39 BPM loss, but it still needs a robust "large" model for motion periods. Our method is a way to find that model. The two approaches are complementary, so we evaluate the selected architecture inside an ActPPG-style switch as well as alone.

Proposed Experiments and Decisive Checks

The primary measurement is Kendall's τ and top-10% regret between each proxy and three targets on PPG-TCN-Bench: overall MAE, motion-activity MAE, and post-transition MAE. Regret is reported overall and within each memory tier. The idea is supported if three things hold together:

  1. Standard proxies show a clear drop in τ from overall MAE to motion and transition MAE.
  2. $A$ shows a smaller drop.
  3. $A$ beats the structural regressor on the motion targets with non-overlapping confidence intervals.

Two secondary checks follow. We test whether the adversarial proxy's skip-connection penalty inverts the ranking of residual TCNs, which we expect it to. We also test whether $A$-selected models keep their advantage on Tri-Spectral and SPC 2015.

The end-to-end test runs a proxy-guided evolutionary search under an STM32WB55 memory cap and compares the selected model's DaLiA MAE, motion MAE, energy, and search cost against our MorphNet+PIT rerun. The idea is refuted if $A$'s advantage disappears once receptive field is controlled for, or if its seed variance exceeds the τ differences it is meant to resolve.