How Much Sensor Energy Are We Spending to Fit Biased Self-Reports?

Wearable stress and sleep models are usually trained and scored against what people say about themselves: a 0–10 stress rating, a sleep-quality questionnaire. Those reports are distorted by recall, mood, and personal scale. One person's "5/10" is another person's "8/10", as the StressID authors note. Separately, green-computing research shows that the last few points of accuracy cost the most energy. This note joins the two observations. When the training labels are biased self-reports, some of that expensive final accuracy may be spent fitting each person's reporting quirks rather than their physiology. If so, the battery a wearable burns on high-fidelity sensing past a certain point is wasted.

We propose to measure this directly and then design around it. Sensor fidelity is controlled by sampling rate, converter resolution, and tolerated circuit noise. We sweep these settings on a stress dataset that carries two kinds of labels: participants' self-ratings and the experimenter-controlled task condition. We trace one accuracy-versus-energy curve for each label type and compare them. We then build an energy-aware sensor tuner that treats the self-report as a biased, person-specific reading of a hidden state. The tuner decides how much sensing to buy against that hidden state, not against the raw rating.

Core Claim

The central claim has two parts. The first is that the curve of accuracy against sensor energy flattens at a much lower energy when accuracy is measured against self-reports than when it is measured against protocol labels. The second is that the accuracy that does keep improving against self-reports at high energy is partly accuracy at predicting who the person is and how they tend to rate, not how stressed they are.

A simple calculation shows why the first part should hold. Suppose a binary self-report flips the underlying state with rate $\rho$, independently of the signals. A model that is correct on the true state with probability $a$ then agrees with the report with probability

$$a_{\text{obs}} = (1-2\rho)\,a + \rho .$$

Every gain in true accuracy therefore shows up shrunk by a factor $(1-2\rho)$. Energy-per-point metrics such as the $kWh/p$ of Portet et al. will then overstate the cost of real gains and misplace the point where one should stop.

The second part of the claim concerns bias that is not independent of the signals. When bias depends on the person, and the person is identifiable from high-resolution physiology, extra sensor fidelity can raise $a_{\text{obs}}$ without raising $a$. This is the more dangerous case, and it is testable.

What would be new is concrete. The sensor-noise tuning of Vlasselaer et al. trades per-sensor noise against power for a tree-augmented naive Bayes classifier (a small probabilistic model that links each feature to the class and to one other feature). It reports at least a 100x sensor power reduction within 10% of baseline accuracy. However, it assumes the class labels of its benchmark datasets (a robot navigation set and three activity-recognition sets) are clean. Nobody in this file has asked how that trade-off moves when the label is a subjective report, or proposed a tuner that accounts for it.

Research Program

The work has three stages.

Stage 1: a measurement study on StressID. StressID has 65 participants and 11 laboratory tasks, with ECG, electrodermal activity, and respiration at 500 Hz, plus 0–10 self-ratings. Each recording carries both a protocol label (stressor task versus rest or relaxation) and the participant's own rating. We define a grid of sensing configurations. The grid varies each channel's sampling rate by decimation, its effective converter resolution by requantization (following the 32-bin, 5-bit emulation in Vlasselaer et al.), its added Gaussian noise $s_i$, and which channels are on.

Energy for each configuration comes from a simple front-end model. Analog power scales as $1/s_i^2$, following the circuit argument in that paper. Sampling and conversion cost is linear in rate times bits. Radio cost is linear in transmitted bits. We report energy in relative units and are explicit that the model is a proxy. This is the same class of limitation the file flags for financial and time proxies such as HULK.

For every configuration we train the same classifier families StressID uses: random forests, support vector machines, and small multilayer perceptrons. We score each against both label types, which yields two Pareto fronts.

To diagnose the person-identity effect, we also train a probe that predicts participant identity from the same features. We then test whether the gains against self-reports at high energy track the probe's identifiability more closely than they track protocol accuracy.

Stage 2: a bias-aware tuner. We extend the noise-tuned Bayesian network so that the class node is a hidden state. Its observed child is the self-report, passed through a person-specific ordinal reporting model: each participant gets their own cut-points on the 0–10 scale and a flip rate. The model is fit with expectation-maximization jointly with the sensor-noise child nodes. The greedy Pareto search over $s_i$ then optimizes predicted accuracy on the hidden state.

Its stopping rule is explicit: stop degrading noise once the expected change in hidden-state accuracy falls below the report-noise floor implied by the fitted $\rho$. Protocol labels are never used for fitting or model selection. They serve only as an external check.

Stage 3: transfer. We apply the same analysis to SWELL-KW and to WESAD. For WESAD we first confirm what condition and self-report annotations the release contains before relying on them, because the file does not describe them. This tells us whether the effect is specific to StressID or general.

Baselines

The naive reference is the full-fidelity configuration, meaning all channels at 500 Hz with no added noise. Its accuracy is the ceiling and its energy is the budget that everything else is measured against. Alongside it we run uniform downsampling and fixed duty cycling. These are the obvious engineering answers, and they are cheap enough to evaluate on the full grid.

The strongest existing method for the energy side is greedy Pareto feature-noise tuning from Vlasselaer et al., run on raw self-report labels as the paper would run it. The reported benchmark is at least 100x sensor power reduction within 10% of baseline accuracy on its UCI datasets. There is no reported result on Lacuna for this method on StressID or any self-report-labelled dataset. For the accuracy side on StressID, the strongest reported numbers come from StressID: a physiological random forest reaches F1 = 0.73 on 2-class stress, and average-rule multimodal fusion reaches F1 = 0.72 on 2-class and 0.63 on 3-class.

All methods share the same subject-wise splits, feature pipeline, energy model, and hyperparameter search budget. The tuner and the baseline get the same number of greedy steps.

Beating the baseline means two things. At matched energy, the configuration our tuner selects must reach higher protocol-label F1 than the configuration raw-label tuning selects. Alternatively, it must match raw-label tuning's protocol-label F1 at significantly lower energy. Both comparisons use paired tests across held-out subjects.

The key ablation removes the person-specific reporting model and keeps a single shared flip rate. This isolates whether individual scale differences, rather than generic label noise, drive the effect. A second ablation replaces our model with per-subject z-scoring of ratings before thresholding.

Why not just…?

Why not just z-score each person's ratings and run ordinary noise tuning? This is the real threat. Standardizing each person's ratings addresses exactly the calibration differences StressID documents. It is included as an ablation, and if it matches the full model we will report that the simple fix suffices for the energy decision. We expect it to fall short in two situations. The first is when a participant's ratings are mostly flat, which leaves nothing to standardize. The second is when bias is asymmetric, for example when stress is underreported for social reasons, a pattern the file attributes to social desirability. Z-scoring shifts and rescales a person's ratings but cannot model either of these. The experiment settles the question per participant.

Why not just train on protocol labels and skip self-reports? In the lab this is possible. Outside it, no protocol exists, and the only label available is what the person reports. The review by Liu et al. also argues that objective sensing contextualizes the subjective account rather than replacing it. Protocol labels are also not ground truth: a mock interview need not stress everyone. We therefore use them as an independent second reference, not as a training target. The claim concerns how the two fronts diverge, not which label is correct.

Why not just use a noise-robust loss on the classifier? Robust losses make the classifier tolerate label noise, but they say nothing about how much sensing to buy. Our question is the location of the energy stopping point. We include a robust-loss variant, trained on each configuration, as an additional comparison inside the raw-label tuning loop.

Why not use uncertainty-driven querying of the user instead? Ragav and Gudur cut labels by 40%, reaching 90.38% versus 91.92% on SWELL-KW. It treats each answer as clean, however, and needs 10 stochastic passes, taking about 5 s against 9 ms on a Raspberry Pi 2. It reduces the burden on the user but does not address report bias or sensor energy, so it complements this work rather than replacing it.

Proposed Experiments

The decisive check is the Stage 1 comparison of the two fronts on StressID. The hypothesis fails if the self-report front and the protocol front flatten at the same energy, and if identity-probe accuracy does not predict the high-energy gains against self-reports. The hypothesis is supported if the self-report front keeps rising while the protocol front has flattened, and the rise tracks identifiability.

The second check is Stage 2. The configuration our tuner selects, which never sees protocol labels, should land closer on the protocol front to the configuration an oracle tuner would pick using protocol labels than the raw-label configuration does, and at lower energy.

The third check is transfer: repeating the fronts on SWELL-KW, and on WESAD once its annotations are confirmed, to see whether the fitted flip rate $\rho$ predicts how early each dataset's self-report front flattens.

We will summarize each front with the harmonic accuracy-energy score of Li et al., computed separately for each label type. As a final sanity check, we will measure a small subset of configurations on a microcontroller development board to verify the rank order predicted by the energy model.