Detection Yield per Charge: Choosing Wearable ECG Pipelines for Long-Horizon Recurrence Surveillance

Clinical studies that predict atrial fibrillation (AF, an irregular heart rhythm that often returns after treatment) usually follow patients for only a few months. Recurrences that appear later are missed. Continuous wearable ECG monitoring could extend that follow-up, but a wearable runs on a battery, and every signal-processing choice trades accuracy against battery life. The algorithms in question include the sampling rate, the heartbeat detector, and whether a neural network classifier runs on the device. Today these algorithms are ranked by per-recording accuracy on short retrospective benchmark strips of 9 to 61 seconds. That ranking ignores the quantity that matters for surveillance: how many true recurrences a device catches before it has to be recharged, removed, or abandoned.

We propose to evaluate and select wearable ECG pipelines by detection yield under a fixed energy budget. This is the expected number of AF episodes detected over a realistic monitoring horizon, and it depends jointly on per-episode sensitivity and on how many hours the pipeline can keep the sensor on. The central claim is that, under realistic battery budgets and episode rates, a cheaper and less accurate pipeline that monitors longer can catch more recurrences than the benchmark leader. The question of whether a retrospective short-strip evaluation predicts which pipeline wins in a forward-enrolled cohort can then be tested directly. The work treats the energy trade-off as the mechanism, limited follow-up as the clinical problem it addresses, retrospective benchmarks as the cheap discovery phase, and a prospective cohort as the decisive test.

Why this gap is real

The motivating evidence comes from both ends of the problem. On the clinical side, Yang et al. built a model from echocardiography and blood BNP to predict AF recurrence after ablation, reaching an AUC of 0.814 in 215 patients, of whom 25.6% recurred. The authors state that their 3–6 month follow-up missed later recurrences. Their outcome is therefore itself a truncated observation, and extending the follow-up is a measurement problem as much as a modeling one.

On the device side, the QRS detection review describes the core tension. Wavelet detectors reach 99.89% sensitivity but are computationally heavy on battery hardware. Derivative-based detectors run roughly 10x faster on phone-class hardware but degrade under motion noise.

The PTB-XL benchmark found that 500 Hz sampling gave no significant diagnostic gain over 100 Hz. That result was obtained on clean 10-second clinical strips, which is exactly the setting where energy savings look free.

The prospective SmartVest study is the constraining evidence. In 10 people over 24 hours it found heart rate and HRV errors under 10–14%, but a PAC error of 61.60% and a PVC error of 395.95%, because motion artifacts mimicked ectopic beats. That study used a fixed pipeline with no adaptive energy control, and it followed patients for one day.

Together these papers say three things: short clean strips flatter cheap pipelines, free-living noise punishes them, and nobody has asked which effect dominates once monitoring duration is part of the objective.

Core Claim

For a device with energy budget $E$ running pipeline $k$ at average power $p_k$, monitoring time is $T_k = E/p_k$. Let $S_k$ be per-episode sensitivity under free-living noise. Suppose AF episodes arrive at rate $\lambda$ and each lasts long enough to be caught when the monitor is on. Then the probability of detecting at least one recurrence in a patient is approximately

$$ Y_k = S_k \left(1 - e^{-\lambda E / p_k}\right). $$

This follows from modeling episodes as a Poisson process observed only while the device is on. $Y_k$ is the detection yield, and false alarms per monitored day $F_k$ act as a constraint.

The formula makes the regime structure explicit. When $\lambda E/p_k$ is large, the bracketed term saturates, the patient will be monitored through an episode regardless, and sensitivity dominates, so accurate pipelines win. When episodes are rare relative to battery life, coverage dominates and cheap pipelines win. What is new is not the formula itself. It is the empirical program that estimates $S_k$ and $p_k$ honestly, tests whether retrospective estimates transfer, and adds adaptive pipelines that sit off this fixed-pipeline curve.

Research Program

We will build a family of on-device pipelines spanning the energy range, and measure the average power of each on a microcontroller-class wearable platform. The family runs from a derivative-based QRS detector with RR-irregularity AF rules at 100 Hz, through wavelet detection, to a compact 1D CNN in the style of DenseECG at 100 and 250 Hz. It also includes duty-cycled variants of each.

The main methodological contribution is an adaptive escalation pipeline. A cheap detector runs continuously, and the expensive classifier wakes only when the cheap stage sees RR irregularity or a signal-quality drop. This borrows the idea of stretching work to fit a budget from DVFS scheduling (Wang et al.). It also borrows the lesson of the hierarchical DRL power manager that a cheap local predictor should gate expensive actions.

The study then proceeds in two phases, modeled on the discovery-then-validation design of Sohal et al., who found a dyssynchrony threshold in 44 patients and confirmed it in 50 forward-enrolled ones.

In the retrospective discovery phase, $S_k$ is estimated on the PhysioNet/CinC 2017 single-lead AF recordings and on MIT-BIH Arrhythmia records with MIT-BIH Noise Stress Test noise added at graded levels. We fit how each pipeline's sensitivity falls with noise, and combine these curves with measured $p_k$ and a range of plausible $\lambda$ to predict each pipeline's yield.

In the prospective phase, post-ablation AF patients wear the patch for multiple weeks, with intermittent reference Holter segments for adjudication. A per-patient randomized crossover of pipelines lets us compare them within patients.

A note on replay validity is needed here. CausalSim showed that replaying traces collected under one policy biases evaluation of another, because the trace depends on the old policy's decisions. In wearable ECG, the cardiac signal does not depend on the processing policy, so downsampling or re-running full-resolution recordings is a valid counterfactual. The retrospective-to-prospective gap here comes from two other sources: the noise distribution shifts from clean strips to free-living wear, and effects such as electrode drift, skin-impedance change, and user removal of the device only show up over weeks. The prospective phase is designed to measure those two sources specifically.

Baselines

The naive reference point is standard episodic clinic follow-up of the kind Yang et al. used: scheduled ECGs at visits within 3–6 months, with no wearable. Any wearable arm must beat its recurrence detection rate, and this baseline fixes the clinical floor.

The second baseline is "benchmark leader, always on." We run the DenseECG-style CNN at full rate until the battery is exhausted, recharge on the same schedule as every other arm, and use no duty cycling. This is the strongest existing method on the accuracy axis. DenseECG reports an F1 of 0.880 on the 2017 PhysioNet/CinC Challenge dataset (DenseECG). For context, PTB-XL CNN ensembles reach a macro AUC of about 0.939, but on a different 12-lead diagnostic task. We found no reported result on Lacuna for any wearable AF pipeline's yield under a battery budget. Beating this baseline means detecting significantly more adjudicated recurrences per patient-month of wear, at a false-alarm rate no higher than a pre-registered ceiling.

The third baseline is the cheapest viable pipeline, a derivative QRS detector with RR rules. It isolates the coverage-only extreme.

The ablation that isolates the proposal's key ingredient compares adaptive escalation against fixed duty cycling at matched average power. If escalation does not beat a duty-cycled CNN that simply sleeps on a fixed schedule, the gating logic adds nothing.

Fairness is enforced across all arms. Every arm uses the same hardware, battery, recharge policy, and adjudication protocol. Every learned component gets the same training data (PhysioNet 2017 plus MIT-BIH, with patient-level splits as PTB-XL requires) and the same hyperparameter search budget.

Why not just…?

Why not just follow PTB-XL and run everything at 100 Hz? PTB-XL's finding comes from clean 10-second clinical strips. The SmartVest results show that free-living motion is exactly where morphology-based detection collapses. The noise-graded retrospective analysis tests directly whether the 100 Hz penalty grows with noise. We include 100 Hz arms, so if the finding transfers, the experiments will show it.

Why not just augment training with Noise Stress Test noise and keep ranking by F1? This is a real threat for the sensitivity term. Augmentation may close much of the noise gap, so we include augmented versions of every pipeline. Augmentation cannot, however, address the coverage term, because F1 has no notion of hours monitored. The claim fails if augmented-F1 ranking and yield ranking agree prospectively. That agreement is the first decisive check.

Why not just skip the retrospective phase and run the prospective study? Prospective adjudicated monitoring is expensive, as the SmartVest study's 10-person, one-day scale illustrates. It also cannot sweep dozens of pipelines and noise levels. The retrospective phase exists to shortlist pipelines, and the prospective phase tells us how far to trust that shortlist.

Why not apply CausalSim-style debiasing? Its target bias is endogeneity of traces to the policy, which is weak here, as argued above. Adding it would add complexity without addressing the actual gap, which is noise shift and duration.

Decisive Checks

Three results decide the claim.

The first is rank concordance, measured by Kendall's tau between the retrospectively predicted yield ordering and the prospectively observed ordering of pipelines. We will also measure it for the F1 ordering. The proposal succeeds if yield ordering transfers significantly better than F1 ordering.

The second is whether the prospectively best pipeline under the energy budget is a cheaper one than the benchmark leader in the low-$\lambda$ regime the formula predicts. This regime can be checked by stratifying patients by observed episode frequency.

The third is whether adaptive escalation lies above the fixed-pipeline yield curve at matched power. If all three fail, the negative result is still useful: it would show that for AF surveillance, per-strip accuracy is an adequate stand-in and energy-aware selection is unnecessary.