Spiking neural networks have been studied for biomedical signals for over a decade, but the practical case for them only became clear once on-chip neuromorphic platforms started appearing [22][23]. Recent work covers ECG arrhythmia detection on Loihi-class hardware [24], EEG analysis with wavelet-front-end SNNs, human-activity recognition for wearables [25], and a 2026 review of low-power cardiac monitoring that surveys the whole space. Two design choices have done most of the heavy lifting: surrogate-gradient training [26][27], which makes it possible to train SNNs with standard PyTorch- or TensorFlow-style backpropagation, and adaptive spike encoding (delta modulation, rate coding, latency coding), which controls how much information is packed into each spike train [28]. NeuroPulse-Edge uses both but goes one step further by also putting the attention block in the spike domain, which is a much less explored design space.

On the other side of the design space, attention has become the default building block for ECG and PPG classifiers [29][30]. A CNN-Transformer hybrid has been shown to be the right combination for real-time ECG anomaly detection, and the foundation-transformer line of work [31] has pushed self-supervised pre-training onto large unlabeled ECG corpora. The downside is energy. A tiny transformer of around 1 M parameters can hit 90%+ accuracy on a five-class arrhythmia task, but the attention block alone takes tens of mW to run continuously [32], which is several times the budget a coin-cell wearable can afford. Spikformer-style architectures resolve this by replacing the dot-product attention with a spike-friendly variant. A hybrid SNN-Transformer that handles both motor imagery and sleep apnea on EEG and ECG has been demonstrated [33], and the 2026 cardiac-SNN review sketches the design template that NeuroPulse-Edge follows. We extend that template by collapsing the attention into a fully binary AND-OR plus popcount, which is what gives us the order-of-magnitude power saving in Section 5.

TinyML has converged on a small set of recipes — pruning, weight clustering, INT8 post-training quantization, and model-architecture search constrained by a target microcontroller [34]. For wearable health specifically, the recent literature includes ECG anomaly detection on Raspberry Pi and Arduino platforms, TinyML-based stress and sleep monitoring on a microcontroller-class AI sensor, an edge-aware IoT framework for real-time health monitoring [35], and a TinyML wearable for HR / SpO₂ / temperature anomaly detection [36]. The reported continuous power numbers cluster between 30 and 100 mW, and the latency between 30 ms and 3 s, which is acceptable for periodic checks but borderline for clinical-grade anomaly detection during exercise. NeuroPulse-Edge aims squarely at the clinical-grade end of that spectrum and pushes the envelope by an order of magnitude in power.

Public wearable benchmarks have improved noticeably in 2024–2026. CACHET-CADB [37] provides 259 days of contextualized ambulatory ECG from 24 patients and remains the most realistic free-living arrhythmia benchmark. WildPPG [38] adds a long, naturalistic outdoor PPG corpus that is roughly eight times the size of PPG-DaLiA. The 2025 Galaxy-Watch PPG release [39] and a 2026 long-term smartwatch arrhythmia dataset [40] cover the consumer-wearable end of the spectrum, and the 2025 multimodal stress dataset [41] adds facial expressions on top of physiological signals. We use four of these datasets and list them in Table 1.

The literature reviewed above leaves four issues unresolved, each of which NeuroPulse-Edge is designed to address.

NeuroPulse-Edge closes these gaps by moving the attention block fully into the binary spike domain, reaching a 1.17 mW continuous budget, deriving an explicit spike-sparsity energy bound, and reporting robustness and held-out generalization across four corpora.

Wearable streams are continuous in time but vary widely in dynamic range — an ECG R-peak is a sharp transient, a PPG signal is smooth, an accelerometer trace is bursty. We use a multi-coding encoder that pairs each modality with the encoding it suits best. ECG is encoded with delta modulation, which fires a positive spike whenever the signal jumps by more than a threshold θΔ and a negative spike for the opposite direction:

s_t^Δ=sign(x_t-x_t-1) ⋅1{|x_t-x_t-1 |>θ_Δ} (1)

PPG is encoded with rate coding over a short window of W samples, mapping the signal amplitude to a spike rate r_t:

(2)

Accelerometer and EDA streams are encoded with latency coding, where larger amplitudes fire earlier within a frame of TF time bins:

(3)

Each encoder runs at the native sampling rate of its sensor (250 Hz for ECG, 100 Hz for PPG, 50 Hz for accelerometer, 4 Hz for skin temperature and EDA), and the four spike trains — s_t^Δ , s_t^r, and

sτ — are then resampled to a common 100 Hz step before they enter the core.

The core is built on the standard leaky integrate-and-fire (LIF) neuron, which approximates a biological neuron with three dynamics — leaky decay, input integration, and a threshold-crossing reset [42]. The membrane potential V_t evolves discretely as:

(4)

where β = exp(−Δt/τ_m) is the leak factor, and τ_m is the membrane time constant. A spike is emitted whenever V_t crosses the firing threshold θ:

ot = 𝟙{Vt ≥ θ} (5)

After firing, the membrane is reset by subtraction (rather than to zero), which preserves the residual energy and improves training stability:

Vt ← Vt − θ ⋅ ot (6)

We stack four LIF layers with hidden width 64 and τ_m = 12 ms, which we found in early experiments to be the smallest configuration that does not lose minor R-peak structure on noisy CACHET-CADB segments.

Between the LIF layers, we use a small spike-conv (S-Conv) block — a 1-D convolution applied directly on binary spike trains, with the convolution itself implemented as integer accumulation followed by a learned threshold:

yt = LIF(Conv1Dw(st)) (7)

We use kernel size 5 and stride 1, with two stacked S-Conv blocks. Because the input is binary, the convolution simplifies to a sum of selected weights — no multiplication is needed at inference time. This is the same trick that makes binary-weight networks fast on microcontrollers, reused here at the activation level.

The piece of work that distinguishes NeuroPulse-Edge from earlier wearable SNNs is the attention block. Following the Spikformer family, we project the LIF output into three binary spike streams

Qs, Ks, Vs ∈ {0,1} T x d:

Qs = LIF(WQS),  Ks = LIF(WKS),  Vs = LIF(WVS) (8)

The attention map is then computed as the popcount of the elementwise AND of the query and key streams, which replaces the dot product without any floating-point MAC:

A_i,j=popcount(Q_s [i,:] AND K_s [j,:]) (9)

The output is the OR of the value rows weighted by a thresholded attention mask A_i, j, fed back into a final LIF neuron to produce h_i :

h_i=LIF(⋁_j (V_s [j,:] AND 1{A_i,j≥τ_a})) (10)

On a Cortex-M4, the AND, OR, and popcount instructions are all single-cycle, which is what gives us the latency advantage in Section 5. We use d = 32 and a small temporal window of T = 32 spikes, which corresponds to roughly 320 ms of biosignal context — long enough to span the full PQRST complex of one heartbeat. The thresholded attention uses τ_a as the activation cutoff.

SNNs are not differentiable at the spike-emission step, so we train with surrogate gradients. The forward pass is the discrete LIF rule; the backward pass replaces the Heaviside derivative with a smooth surrogate σ, ′ such as the box function or a fast sigmoid:

(11)

with surrogate width γ. The loss is a standard cross-entropy on the temporally pooled output of the final LIF layer, summed over the T = 32 spike steps:

(12)

This subsection establishes two properties of NeuroPulse-Edge: a convergence guarantee for surrogate-gradient training, and an energy bound that ties inference cost to spike sparsity.

Once the floating-point model is trained, we apply post-training INT8 quantization to the residual continuous parameters — namely the projection matrices W_Q,WK, WV, Wo, and the membrane decay factor β. The classifier head is a single linear layer with INT8 weights and INT32 accumulators, run through a softmax-free arg-max to produce the prediction ŷ:

(16)

The complete model is around 96 KB on disk and runs on a 256 KB Flash, 64 KB RAM Cortex-M4 testbench. We do not retrain after quantization; the spike-driven nature of the activations makes the model relatively quantization-tolerant, which is consistent with what the broader SNN literature reports.

Algorithm 1 summarizes the full inference pipeline. The encoder, LIF stack, S-Conv blocks, spike attention, and INT8 classifier are all run in sequence; the wake-up gate runs in parallel and can short-circuit the radio when the spike rate is below a quiet-period threshold ρ_quiet.

Algorithm 1. NeuroPulse-Edge inference loop

Input: streaming sensor frame (xECG, xPPG, xACC, xTEMP)

Output: on-device decision ŷ

for each incoming sensor frame do

sECG ← deltaMod(xECG)

sPPG ← rateCode(xPPG)

sACC, sTEMP ← latencyCode(xACC, xTEMP)

S ← alignTo100Hz([sECG, sPPG, sACC, sTEMP])

if meanSpikeRate(S) < rhoQuiet then

return SLEEP

end if

H ← LIFstack(S)

H ← SConv(H)

H ← SAttn(H)

ŷ ← INT8classify(poolT(H))

if ŷ in {AF, PVC, PAC, Other} then

transmit BLE alert

end if

return ŷ

We evaluate NeuroPulse-Edge on four wearable benchmarks chosen to cover the realistic operating conditions of a 2026 wearable. CACHET-CADB provides 259 days of contextualized ambulatory ECG from 24 patients and 1,602 manually annotated 10-second heart-rhythm samples; we use it as the primary arrhythmia benchmark. WildPPG is a long, real-world outdoor PPG corpus released in late 2024 / early 2025, used for HR / HRV estimation and noise-robustness tests. The 2025 Galaxy-Watch PPG release adds a semi-naturalistic indoor/outdoor PPG corpus collected on a consumer smartwatch. The 2025 multimodal stress dataset supplies the EDA + skin-temperature + facial-expression streams for the stress-state head. Finally, a 2026 long-term smartwatch arrhythmia release is used as a held-out generalization test. Table 2 summarizes the five datasets.

All four corpora are placed on a uniform processing track before training. Per-modality preprocessing proceeds in three stages. (i) Each raw stream is band-pass filtered to its physiological band (0.5–40 Hz for ECG, 0.5–8 Hz for PPG, 0.1–20 Hz for accelerometer and EDA) and resampled to its native rate. (ii) Amplitudes are normalized per recording using robust min–max scaling computed on the training partition only, so that no information from validation or test segments leaks into the scaling parameters. (iii) The normalized streams are encoded into spike trains by the multi-coding encoder of Section 3.1 and aligned to a common 100 Hz step.

Records with more than 50% missing samples in any channel are dropped; shorter gaps are zero-filled at the spike level (no spike), which is the natural missing-data behavior for an event-driven encoder, and a per-channel availability flag is carried alongside the spike train. Each corpus is partitioned 70/15/15 (train/validation/test) by subject, so that no subject appears in more than one partition and the test accuracy reflects genuine inter-subject generalization. The 2026 smartwatch arrhythmia release is held out entirely and used only for the final generalization test. Five-fold subject-stratified cross-validation is applied within the train/validation portion. All reported numbers are means with sample standard deviations over five independent random seeds, {7, 19, 23, 42, 71}, used for both weight initialization and mini-batch ordering. Training uses the Adam optimizer (β₁ = 0.9, β₂ = 0.999) with a cosine-decayed learning rate starting at 1×10⁻³ and decaying to 1×10⁻⁵ over 150 epochs, batch size 128, and surrogate-gradient width as listed in Table 3. Early stopping is applied with a patience of 15 epochs monitored on the validation macro F1; the best-performing checkpoint is retained and used for all test-set evaluations.

All on-device numbers are measured on an STM32L4R5ZI ARM Cortex-M4 board running at 80 MHz with 256 KB Flash and 640 KB SRAM, which is representative of the silicon used in current-generation health-grade wearables. Power is measured with an INA219 shunt monitor, following the protocol. Latency is measured at batch size 1 over 10,000 frames. The SNN backend is SpikingJelly for training and a hand-rolled C-with-CMSIS-NN runtime for inference. We compare NeuroPulse-Edge against five baselines: (i) MobileNet-V2 with INT8 quantization, (ii) a 1-D CNN tuned for ECG, (iii) a tiny transformer of around 0.6 M parameters, (iv) a bidirectional LSTM, and (v) a plain SNN without the spiking-attention block. All baselines are run on the same testbench with the same preprocessing. Table 3 lists the hyperparameters.

We report four families of metrics. (a) Clinical accuracy: macro accuracy and macro F1 on a five-class arrhythmia task (NSR, AF, PVC, PAC, Other), evaluated on the held-out CACHET-CADB test split. (b) Heart-rate accuracy: mean absolute error (MAE) of the on-device HR estimate against the ECG reference, computed on WildPPG. (c) Energy: continuous power draw in mW, measured on the STM32L4R5ZI testbench at full duty cycle. (d) Latency: end-to-end per-inference time in ms, also on the testbench. All numbers are averaged over the five random seeds; the per-component power breakdown is averaged over 10 minutes of free-living recording. Ninety-five percent confidence intervals (95% CIs) are derived from the reported standard deviations using the t-distribution with four degrees of freedom (t₀.₁₂₁, 4 = 2.776), following the formula 95% CI = mean ± 2.776 × (SD/√5). Representative CIs for the primary result are: NeuroPulse-Edge accuracy 95% CI [91.73%, 92.47%], macro F1 95% CI [0.916, 0.926], and continuous power 95% CI [1.07 mW, 1.27 mW].

Differences between NeuroPulse-Edge and each baseline are assessed for statistical significance with the two-sided paired Wilcoxon signed-rank test over the matched per-seed, per-fold results. Because the same baseline set is compared on several metrics and datasets, the Holm–Bonferroni step-down procedure is applied to control the family-wise error rate across the resulting family of comparisons. The Wilcoxon test is preferred over a paired t-test because the normality of fold-level differences cannot be credibly assumed from five observations. Every improvement reported below clears the Holm-corrected p < 0.01 threshold.

Figure 3(a) reports the surrogate-gradient training curves on the joint CACHET-CADB + WildPPG mix. The model converges in roughly 100 epochs, and the validation accuracy plateaus at 92.5%, which is in the same range as recent transformer-based ECG classifiers but at a fraction of the cost, consistent with the convergence behavior predicted. Figure 3(b) reports a component-level ablation on the five-class arrhythmia task. Removing the spiking-attention block (i.e., running plain LIF + S-Conv) costs 7.0 macro-F1 points; replacing the LIF neurons with ReLU activations costs 9.2 points and breaks the energy story entirely; removing the delta-modulation encoder and using a fixed rate-coding scheme everywhere costs 5.7 points. The FP32 (no-quantization) baseline lands 0.7 points above the full INT8 model — a small accuracy headroom we trade for a roughly 4× memory and a 2× latency reduction, which is the kind of trade most wearable engineers would happily take.

Figure 4(a) reports the per-component power breakdown of the deployed model on the STM32L4R5ZI testbench. The total continuous draw is 1.17 mW, of which the sensor front-end (ECG and PPG analog stages) accounts for 0.42 mW, the radio for 0.32 mW, and the entire neural network — encoder, LIF, S-Conv, S-Attn, and the INT8 head — for only 0.43 mW. The S-Attn block costs 0.11 mW, which is roughly an order of magnitude cheaper than the equivalent floating-point attention in the tiny-transformer baseline, in line with the spike-sparsity bound. The trick is the AND-OR-popcount substitution on a Cortex-M4; those three primitives are single-cycle and dispatch into the existing register file without ever touching the floating-point unit.

Figure 4(b) shows the confusion matrix on the five-class arrhythmia task. The diagonal stays above 0.88 in every class, with the most common confusion (PAC versus PVC, 0.05) being a known cardiology hard case rather than a model artifact. NSR is the easiest class at 0.96, AF lands at 0.93, and the residual error on the rare classes (PVC, PAC, Other) is dominated by motion-induced PPG noise rather than by ECG ambiguity, which is what one would expect from an ambulatory free-living dataset.

Figure 5 and Table 4 compare NeuroPulse-Edge against the five baselines on accuracy, continuous power, and latency. NeuroPulse-Edge wins on all three axes. On accuracy, it lands at 92.1%, narrowly beating the tiny transformer (91.2%) and clearly beating the plain SNN (86.5%). On continuous power, it draws 1.17 mW versus 12.8 mW for the tiny transformer and 6.4 mW for the 1-D CNN, an 8.2× and 5.5× reduction, respectively. On latency, it answers in 14 ms versus 84 ms and 41 ms, a 6× and 3× reduction. The accuracy gap to the tiny transformer (0.9 points) is the price the model pays for being event-driven, and the power and latency gaps are what it earns in return — a trade that is clearly worth it on a coin-cell budget. All pairwise differences are clear, p < 0.01 (Wilcoxon signed-rank with Holm–Bonferroni correction).

Table 5 reports the per-dataset accuracy of the four strongest models, including a held-out generalization test on the 2026 smartwatch arrhythmia dataset. The pattern is consistent: NeuroPulse-Edge is the best on three out of four datasets and is within 0.3 points of the tiny transformer on the fourth (multimodal stress, where attention-rich models have a small edge). The held-out 2026 release is the most informative number — it is the only dataset that the model was not tuned on at all, and NeuroPulse-Edge still generalizes to 88.6% accuracy, which is in the range of clinical-grade thresholds reported by recent FDA-cleared algorithms.

Two further experiments probe practical reliability and deployment reach. The first examines robustness to imperfect inputs on the CACHET-CADB test set under three perturbations: additive Gaussian noise at 15 dB SNR, simulated motion artifact injected as band-limited baseline wander at 20% of signal amplitude, and a random 10% channel dropout (one of the four modalities removed at random per window). Table 6 reports the resulting macro F1. NeuroPulse-Edge degrades the least in every condition, which is consistent with the event-driven encoder discarding sub-threshold noise before it reaches the core and with the redundancy of the multimodal spike streams.

The second experiment examines scalability across microcontroller classes, since absolute power and latency depend on the deployment target. The full model is cross-compiled to four representative cores, and Table 7 reports continuous power, per-inference latency, and the resulting accuracy (which is invariant, since the network is identical). The spike-only design keeps the model within a sub-2 mW envelope from the Cortex-M0+ up to the Helium-equipped Cortex-M55, with latency scaling roughly inversely with clock and SIMD width.

To assess the robustness of the proposed architecture to the choice of spike-encoding hyperparameters, we vary four key parameters of the multi-coding encoder — the delta-modulation threshold θΔ, the rate-coding window width W, the latency-coding frame length TF, and the maximum spike rate rmax — independently around their default values while keeping all other settings fixed. Table 8 summarises the resulting accuracy, macro F1, and continuous power on the CACHET-CADB test set (mean ± SD over five seeds).

Two trends are evident. First, the delta-modulation threshold θΔ has the largest influence: halving it to 0.025 mV causes a 0.7-point accuracy drop as sub-threshold noise generates excess spikes, whereas increasing it beyond 0.063 mV begins to suppress genuine R-peak transitions. Second, the rate-coding window W and latency frame TF show roughly symmetric degradation around their defaults, confirming that the chosen values sit near a local optimum. Across all 14 configurations in Table 7, accuracy varies by at most 1.3 percentage points and macro F1 by at most 0.016, which demonstrates that NeuroPulse-Edge is not critically sensitive to the exact encoder configuration and that the defaults transfer well to the tested parameter range.

Three takeaways come out of the numbers. First, the spiking-attention block does most of the work. The 7.0-point F1 drop when it is removed is a much larger effect than any other ablation, which means the temporal-attention idea — the core insight of the transformer literature — survives the move into the spike domain. Second, the power story is real. The 1.17 mW continuous draw on a Cortex-M4 testbench is in the range that lets a coin-cell wearable run for weeks rather than hours, and most of that draw comes from the analog front-end and the radio rather than from the neural network itself, exactly as the spike-sparsity bound predicts. Third, accuracy is not a zero-sum trade-off with energy here: the full model is the best on accuracy and the best on energy at the same time, which is the realistic expectation for a well-designed event-driven system rather than a surprise. The catch — there is always one — is that this only holds when the input has the kind of temporal structure that LIF neurons exploit, namely sharp transients on top of a slow baseline; the same architecture would not buy nearly as much on, say, a stationary tabular medical record.

Several limitations should be stated explicitly. (i) The hardware evaluation centers on a single board (STM32L4R5ZI), and although Section 5.5 cross-compiles to four cores, absolute power numbers on substantially different silicon will shift; the relative ordering of the methods is expected to hold. (ii) The clinical evaluation uses five arrhythmia classes; a 17-class evaluation, closer to a true Holter-style workload, would expose more inter-class confusion. (iii) The wake-up gate assumes that quiet periods of the input correspond to absence-of-anomaly periods, which holds for arrhythmia but not necessarily for slowly emerging conditions such as ischemia, where the model might wake up too late. (iv) The privacy story is incomplete — on-device inference removes much cloud-side risk, but a serious deployment must still reason about side-channel and model-extraction attacks, and recent generative-AI work on synthetic biosignals suggests an obvious next step for both augmentation and privacy. (v) Surrogate-gradient training of SNNs remains less stable than backpropagation on a standard CNN, and a careful learning-rate schedule was needed to converge reliably; adaptive surrogate functions could simplify this.