Hospitals that want to measure staff stress continuously rely on two things. Workers wear sensors, and fixed radio nodes around the building collect those sensors' data and locate the wearer. A promising way to run such nodes without batteries or wall outlets is to power them wirelessly from a dedicated radio transmitter. The physics then imposes a strong rule. A circuit that turns radio waves into usable power needs a signal roughly a thousand to a million times stronger than a data receiver does, so self-powered relays must sit close to the power source. Our observation is that this rule turns into a measurement problem. If coverage clusters around transmitters, which in practice end up at nursing stations and workstations, then the stress data that reaches the hospital is drawn mostly from the places and the people the infrastructure happens to favor. ICU nurses spend more than half of their communication time inside patient rooms, and back-office staff are rarely instrumented at all. Data that goes missing in exactly those places is missing for a reason, so averaging over what remains does not cancel the gap; it skews the stress estimates the institution uses to decide where to send support.
We propose to treat relay placement as a measurement-design problem. The goal is to choose where power transmitters and self-powered relays go so that the resulting stress estimates stay accurate and fair across units, shifts, and staff groups, with the radio's power threshold as a hard constraint. Data throughput becomes a secondary concern. The study also delivers a diagnostic that applies to any sensing setup with limited coverage, including the convenience-placed battery beacons used in existing hospital studies. It measures how much infrastructure geometry alone distorts the psychometric conclusions an institution would draw.
Core Claim
The relay-placement literature optimizes how many bits get through. Wirelessly powered relaying work (Optimum Wirelessly Powered Relaying) derives closed-form time-splitting between harvesting and transmitting and shows that the relay's best position follows the power beacon. Related analyses (Liu et al., Cui et al.) show that energy-harvesting relays must move toward the source, and move further still under non-ideal amplifiers. Active reconfigurable surfaces (Transmission Design for Active RIS-Aided Simultaneous Wireless Information and Power Transfer) are best placed near the energy receivers. All of these treat data as interchangeable bits.
Hospital sensing studies sit on the other side of the gap. Feng and Narayanan (2024) placed Bluetooth beacons at workstations and units to track 43 ICU residents. Feng and Narayanan (2022) put hubs in nursing stations, break rooms, and patient rooms for 99 nurses. Both papers treat placement as an empirical coverage exercise and do not ask how coverage gaps bias the stress models trained on what was captured.
Our claim is that placement chosen to maximize throughput under a harvesting constraint produces stress estimates that are measurably biased by location and by staff group. A placement objective weighted by where and when informative stress episodes happen can recover most of the accuracy and group fairness of full coverage while still respecting the power threshold. Formally, we choose positions $P$ for power beacons and relays to solve
$$\max_{P}\; \sum_{x} w(x)\, \Pr[\text{link}(x;P)] \quad \text{s.t. } \; p_{\text{rx}}(r;P) \ge \theta \;\; \text{for every relay } r \in P,$$
where $x$ ranges over locations, $w(x)$ is how much a sensing window at $x$ improves the stress estimate, $\Pr[\text{link}(x;P)]$ is the probability that a wearable at $x$ reaches a powered relay, $p_{\text{rx}}(r;P)$ is the power the relay receives, and $\theta$ is the harvesting threshold (between −20 and −10 dBm per the file).
Prior work never estimates the weight $w(x)$. We learn it from labeled wearable traces. It combines how much time staff spend at $x$, the stress label rate there, and how uncertain the stress model is about windows recorded there. The uncertainty measure is the Monte-Carlo-dropout estimate of Bayesian Active Learning for Wearable Stress and Affect Detection, which runs dropout repeatedly at test time to gauge how confident a small network is.
Research Program
The work has three parts.
Coverage-bias simulator. We first build a simulator driven by real movement traces. It replays staff location and wearable streams from the TILES hospital cohorts, using the hub-level location logs and stress labels described by Feng and Narayanan. A radio layer decides, for each sensing window, whether the window would have reached a powered relay under a candidate placement. That layer uses indoor path loss, attenuation from human bodies, and the harvesting threshold. We will request access to the TILES data. If that fails, we fall back on synthetic movement traces calibrated to the reported time shares, with stress labels from SWELL-KW.
Placement solver. We extend the closed-form harvest-and-transmit time split of Optimum Wirelessly Powered Relaying. We supply the method from that paper but change its objective to the weighted coverage above. We then solve the placement over a floor plan with greedy submodular selection, which fits naturally because weighted coverage has diminishing returns, followed by local refinement.
Institutional decision layer. Following Workplace factors associated with mental health of healthcare workers during the COVID-19 pandemic, support needs differ sharply by role and setting. That paper reports nurses versus doctors at OR 1.64, ICU redeployment at OR 1.67, and poor workplace support at OR 2.49. We therefore assess placements by whether they preserve the unit-by-shift ranking of estimated stress. That ranking is what an institution would use to direct staffing relief or rest time.
The support-asymmetry finding of Fernandes et al. shapes the evaluation. Technical and administrative staff are routinely left out of wellness infrastructure, so we report group-level error separately for peripheral staff zones.
The finding of Does Techno-invasion Trigger Job Anxiety? constrains the design. Technology intrusion strongly predicted job anxiety (β = 0.675), so the system must not become a new stressor. We therefore cap node count, exclude raw audio, and in a small pilot measure techno-invasion and perceived organizational support with that paper's validated scales.
Baselines
The naive reference point is placement by midpoint geometry, the classic rule for battery-powered relays. It ignores harvesting, so many relays end up below threshold and inactive. The cited theory implies this baseline will lose most coverage; we include it to quantify that loss.
The strongest existing physical-layer method is throughput-optimal source-proximal placement with beacon-tracking time splitting, following Optimum Wirelessly Powered Relaying and Liu et al.. It is the right competitor because it is exactly what a wireless engineer would deploy today. For reference, the active-surface system of Ren et al. reports a weighted sum rate of about 60 bps/Hz versus about 45 for passive surfaces under a 30 W budget. For the beacon-relay design itself, no reported numerical throughput result was found on Lacuna. All placement baselines use the same floor plan, node budget, transmit power, channel model, and solver iteration budget.
The strongest existing stress-inference method is the Random Forest on passive wearable and proximity data from Feng and Narayanan (2024). It reports a macro-F1 of 56.20 for midday stress and 63.41 for end-of-day stress. The nurse cohort (Feng and Narayanan 2022) reports a negative-affect F1 of 65.68%. We retrain this model, and the Bayesian network of Ragav and Gudur, on each placement's surviving windows, with identical hyperparameter search. The full-coverage version, equivalent to unconstrained battery or grid hubs, is our upper bound.
Two ablations isolate the key ingredient. In the first, $w(x)$ is occupancy time only. In the second, $w(x)$ is uniform over the floor plan. If occupancy alone matches the full weight, then the learned stress-informativeness adds nothing, and we will say so.
Beating the strongest baseline means three things at the same node budget and power threshold:
- A statistically significant gain in macro-F1 over throughput-optimal placement.
- Lower error in group-level mean stress for patient-room-heavy and peripheral staff.
- A unit-by-shift stress ranking closer to the full-coverage ranking, measured by Kendall's tau.
Why not just…?
Why not just install battery or wall-powered beacons everywhere? This is the most serious threat, and in many hospitals it may be the right engineering answer. The proposal survives it in two ways. First, the bias diagnostic applies equally to battery beacons placed for convenience, as in the TILES deployments. Measuring how placement alone shifts F1 and group means is useful whether or not harvesting is ever adopted. Second, ICU outlets are reserved for life-support equipment and staff cannot service tags inside sterile fields. The question is therefore whether harvesting-constrained placement can come within a small margin of the full-coverage upper bound. If it cannot, we will report that harvested sensing is not yet viable for this purpose.
Why not just impute the missing windows statistically? Imputation works when data is missing at random. Our premise is that it is not: windows go missing in patient rooms, where stress is concentrated. We test this directly by running multiple imputation on throughput-optimal placements and checking whether the bias in group means persists. If imputation closes the gap, the placement objective is unnecessary and the paper becomes a negative result.
Why not just place relays where people spend the most time? This is the occupancy-only ablation. It is a real competitor, because time share and stress may coincide. Night-shift nurses, for example, showed elevated negative vocal arousal at nursing stations (Feng and Narayanan 2022), a high-traffic location. The experiments settle this: if occupancy weighting matches the learned weight within confidence intervals, we adopt the simpler rule.
Why not just ask people with more frequent surveys? Dense momentary surveys tire respondents and reduce compliance. Ragav and Gudur show that uncertainty-triggered labeling reaches 90.38% accuracy with 40% fewer labels, versus 91.92% with all labels, on SWELL-KW. We incorporate this as complementary rather than competing: surveys are prompted only when a powered relay link exists and the model is uncertain.
Proposed Experiments
The first decisive check is whether bias exists at all. On the TILES traces, we compare group-level mean stress and macro-F1 under throughput-optimal placement against full coverage. A gap within noise would end the project's premise.
The second check sweeps the harvesting threshold from −20 to −10 dBm and the node budget. For each setting we trace how much of the full-coverage F1 and ranking fidelity each placement retains.
The third check stress-tests robustness by adding body-shadowing fading derived from staff density in the traces. This addresses the perfect-channel-knowledge assumption of the cited radio formulations.
The fourth check is a small bench pilot in one ward-like space. A dedicated power beacon and a handful of harvesting relays will confirm activation rates against the simulator's predictions. Paired surveys on techno-invasion and perceived organizational support will check that instrumentation does not itself raise anxiety.
The proposal succeeds only if weighted placement beats both throughput-optimal and occupancy-only placement on group-level error and ranking fidelity at equal budget.
References
- Optimum Wirelessly Powered Relaying
- Noncoherent Relaying in Energy Harvesting Communication Systems
- Optimal Energy-Efficient Relay Deployment for the Bidirectional Relay Transmission Schemes
- Transmission Design for Active RIS-Aided Simultaneous Wireless Information and Power Transfer
- Understanding Stress, Burnout, and Behavioral Patterns in Medical Residents Using Large-scale Longitudinal Wearable Recordings
- Exploring Workplace Behaviors through Speaking Patterns using Large-scale Multimodal Wearable Recordings: A Study of Healthcare Providers
- Bayesian Active Learning for Wearable Stress and Affect Detection
- Workplace factors associated with mental health of healthcare workers during the COVID-19 pandemic: an international cross-sectional study
- Promoção da saúde mental de servidores efetivos de universidades federais: uma revisão integrativa da literatura
- Does Techno-invasion Trigger Job Anxiety? Moderating Effects of Computer Self-efficacy and Perceived Organizational Support