Energy per Diagnosis: Spiking Resonator Models for Always-On Cough Screening

An earbud that listens all day could catch coughs and flag possible respiratory disease. But the models that read diagnostic signal from cough audio are dense neural networks, such as the respiratory foundation models in OPERA, and they burn energy on every audio frame. Neuromorphic computing offers another option. Spiking neural networks (SNNs) are networks whose units send sparse binary events instead of continuous values, so they compute only when something happens. They replace multiplications with additions. The research file shows that no one has tested SNNs on any standard respiratory benchmark, whether COUGHVID, Coswara/DiCOVA, ICBHI, or CIDRZ.

The obvious way to fill that gap would be to train an SNN on a cough dataset and report "X% fewer operations." That result would not show anything. Cough benchmarks are pre-cut clips, so they hide the one thing that decides real energy use: how often the device is processing silence, speech, or a cough. A skeptic would also point out that a simple loudness gate in front of a small quantized CNN already removes most of the idle cost. This proposal takes that objection head on. We build a streaming version of existing cough datasets with realistic event rates. We then ask a narrower question: once a cough is happening, can a spiking model with resonator neurons, deciding before the cough ends, reach diagnostic accuracy matched to the best gated, quantized conventional pipeline while using fewer operations per diagnosis? Under this framing, a clear negative result is also worth publishing.

Core Claim

A gate can remove the silence between coughs. It cannot remove the work done during a cough, and that is where an event-driven model can still save energy. Our hypothesis has three parts. First, a learned bank of resonate-and-fire neurons (spiking units that behave like tuned band-pass oscillators) produces events mostly during the explosive and early decay phases of a cough. Second, a small spiking state-space classifier reading those events can commit to a diagnostic decision partway through the cough. Third, together these keep accuracy within one AUC point of a matched non-spiking model and cut total operations per diagnosed cough by a large factor relative to a gated, quantized CNN. "Total operations" means synaptic events plus any remaining multiplies. The new elements are three: the streaming benchmark, the energy-per-diagnosis metric, and the within-cough early-commit mechanism, tested against the strongest conventional design rather than an always-on strawman.

Prior Work and Its Role

The method comes from neuromorphic audio work. Efficient Neuromorphic Signal Processing with Loihi 2 shows resonate-and-fire neurons approximating a short-time Fourier transform with 47x less output bandwidth, trained end to end with surrogate gradients. Surrogate gradients replace the non-differentiable spike step with a smooth stand-in during backpropagation. That paper, however, evaluates only spoken digits and commands. SH²RFSSM turns harmonic resonate-and-fire dynamics into a stable spiking state-space model without GeLU-type layers. It reports about 69x better estimated energy efficiency than the non-spiking LinOSS on EigenWorms, plus continuous regression on PPG-DaLiA. We reuse its architecture as the classifier and regressor. Two other papers motivate the early-decision and noise-robustness angles. SOM-SNN decides correctly after hearing 50% of an environmental sound. SATRN reaches 95.5% on UrbanSound8K against 95.9% for an ANN, with about 80% fewer operations. Neither paper touches respiratory audio, and neither faces a gated baseline.

Several papers constrain the design. HearCough runs a 12.20 MFLOP, 385 kB CNN on a Cortex-M4F and reaches 95.4% accuracy at detecting the wearer's own coughs at 8 kHz. Tiny conventional detection is therefore already solved. Our contribution has to lie in diagnosis, and we adopt its 8 kHz finding. Patient-specific tuning on ICBHI shows that log quantization cuts memory 4x without significant loss, so quantization is a baseline we must include. Spiking Cochlea with AGC warns that analog spiking front-ends saturate on loud transients and that its 0.5 ms gain loop is untested on explosive sounds. For that reason we use a digital, learned resonator bank and make no claims about silicon cochleas. On the diagnostic side, the KK pediatric cough study shows that fine-grained pathology classification is hard: 47.9% at the cough level for 4 classes. The cough regression benchmark finds that frozen foundation models barely beat a mean predictor on several CIDRZ targets. These results bound what any efficient model can hope for. We report accuracy relative to these ceilings, not absolute claims.

Research Program

The first step is the dataset. We build a streaming cough benchmark by inserting labeled cough clips from Coswara (with DiCOVA's development and evaluation subject splits), COUGHVID, and CIDRZ into hours of continuous background audio. The background contains speech, laughter, eating, walking noise, and ambient scenes, at controlled signal-to-noise ratios and controlled cough rates. Rates run from a few coughs per day to dozens per hour. The streams are built so that the subject splits stay disjoint. The benchmark reports energy per hour of listening and energy per diagnostic decision as functions of cough rate. This turns cough datasets, which today measure only clip accuracy, into a setting where efficiency claims can be tested.

The second step is the model. The front-end is a bank of 32 to 64 learnable resonate-and-fire neurons on raw 8 kHz audio, with frequencies initialized to cover 100 Hz to 4 kHz. Its events feed a three-stage spiking pipeline. The first stage is a detection head. The second is an SH²RFSSM diagnostic classifier. The third is an early-commit rule that outputs a decision once the running class margin passes a threshold tuned on validation data. A spike-rate penalty in the loss lets us trace the full accuracy-versus-events curve.

The third step is accounting. We count synaptic accumulates, residual multiplies, and memory reads separately for every model. We report the raw counts and also energy estimates under per-operation energy constants that we state explicitly. Where access allows, we deploy the network on Loihi 2 to check that the estimates match measured energy.

Baselines

The naive reference points are the class prior or mean predictor (CIDRZ regression) and an always-on dense CNN with no gate. The first sets a floor for accuracy. The second sets a ceiling for energy.

The published diagnostic reference is the DiCOVA BiLSTM on log-Mel features, which the Second DiCOVA Challenge reports at 74.9% AUC on the cough track. We rerun it on our splits. The accuracy ceiling is frozen OPERA-CT and OPERA-GT embeddings with the MLP-small head, which the regression benchmark found best in 23 of 30 cases. We found no reported result on Lacuna for OPERA on the DiCOVA cough-track split, so we will measure it ourselves. For CIDRZ regression, we report results as improvement over the mean predictor, because the file says only that models beat it by 1 to 8%.

The strongest competitor is a gated, quantized conventional pipeline. It has an energy gate, a HearCough-style detector, a quantized diagnostic CNN distilled from OPERA, and its own early exit on truncated input. It gets the same data, the same 8 kHz input, and the same hyperparameter search budget as the SNN, measured in GPU-hours and number of trials. It is the strongest competitor because it captures every efficiency gain that does not require spikes. Beating it means two things at once. AUC must not be significantly worse, defined as within one point under a paired subject-level bootstrap. Operations per diagnosed cough must be at least 3x lower, at both low and high cough rates.

Two ablations isolate the key ingredients. The first is an identical non-spiking oscillatory state-space model of the same size, a LinOSS-type model. It isolates the effect of spiking itself. The second swaps the resonator front-end for LIF neurons and for a Mel front-end, and turns early commit off. This shows which parts produce the savings.

Why not just…?

Why not just quantize a standard CNN? Quantization shrinks memory, 4x on ICBHI in the patient-tuning study, but every frame still pays for dense multiplies. That argument alone is not enough, which is why quantization is built into our strongest baseline instead of being argued away. The streaming benchmark will settle the question. If quantized ANNs match the SNN's energy at equal AUC, the neuromorphic case fails here.

Why not just gate a tiny CNN on loudness? This is the real threat. Between coughs, a gate captures most of the idle savings an SNN would offer. Our claim rests only on within-cough sparsity and early commit, and the cough-rate sweep tests it directly. We expect that at very low cough rates the gate's fixed cost dominates and both systems converge. We will report that crossover point honestly.

Why not just send detected coughs to a phone running OPERA? Offloading keeps accuracy at the ceiling, and HearCough-class detectors make it feasible. We include this variant and report bits transmitted per cough. Our operation counts do not include radio energy, so we treat offloading as a separate regime. Our proposal matters only where offline, private, on-device diagnosis is required.

Decisive Checks

The central experiment plots AUC on the DiCOVA cough-track subjects and on COUGHVID, together with improvement over the mean on CIDRZ age and BMI, against operations per diagnosed cough. It does this for every model at three cough rates and three signal-to-noise ratios. The hypothesis is supported if the spiking resonator model's curve lies to the left of the gated quantized pipeline's curve with no significant AUC loss. A second check measures how much of each cough the early-commit rule actually uses. If decisions need more than 80% of the cough, early commit contributes little. A third check looks inside the trained network: are the learned resonator frequencies concentrated in the 350 Hz to 4 kHz band of the explosive phase, and do event counts per cough track energy rather than pathology? If they track energy, the savings would vanish in loud environments. If the gated baseline matches the SNN within 2x, we will report that conventional gating is sufficient for cough screening. Neuromorphic work has lacked that kind of benchmark-grounded comparison.