When One Armchair Blinds Every Sensor: Correlated Blockage and the Real Follow-Up of Passive Health Monitoring

Trials of home-based interventions for older adults keep ending at the same question: did the benefit last? Studies of digital cognitive training and virtual-reality exercise report gains at the end of the program but cannot say whether those gains persist. The obvious fix is to keep watching patients cheaply after the trial ends using passive sensors, such as millimeter-wave radar, a short-wavelength radio sensor that tracks body motion without cameras or anything the patient has to wear. These sensors only see a person along an unobstructed straight path. Furniture, walls, and other people block that path. Planners usually assume each sensor is blocked independently of the others, so adding sensors appears to shrink blind time multiplicatively. That assumption is wrong. A single sofa or a second resident can block several sensors at once. Worse, those shared blind spots are not random. They are the places where a frail person sits, rests, or moves least.

This note proposes to treat spatially correlated blockage as the mechanism that sets the effective follow-up of a passive-monitoring study. The claim has two parts. First, the independence assumption overstates how many observed person-hours a multi-sensor home deployment will deliver, so studies are planned too short or with too few sensors. Second, the hours that are lost are lost informatively, which biases estimates of how mobility changes over months. We would build a geometric model of indoor obstacles that predicts both effects before deployment, and use its predicted observation probabilities to correct the estimated trajectories.

Core Claim

Let $T$ be the nominal follow-up time, $K$ the number of sensors, and $B_k(x)$ the event that sensor $k$ is blocked when the patient stands at location $x$. A person at $x$ goes unobserved only if every sensor is blocked there. The observed fraction of follow-up time is therefore an average over where the patient actually spends time, $\pi(x)$:

$$T_{\text{eff}} = T \int \pi(x)\,\big[1 - P(B_1(x)\cap \dots \cap B_K(x))\big]\,dx .$$

Under the independence assumption the joint term becomes $\prod_k P(B_k(x))$. When one obstacle cuts several sight lines, that product is far smaller than the true joint probability. This is the same overestimation of coverage that Impact of Blocking Correlation on the Performance of mmWave Cellular Networks proved for cellular links. The second part of the claim is that $\pi(x)$ and the outcome are linked. Low-mobility behavior concentrates in furniture-dense zones, where joint blockage is highest. As a result, observed data over-represent active periods. A naive estimate of month-over-month decline in mobility is then biased toward "no decline," which is exactly the durability question the trials could not answer.

Where This Sits in Prior Work

The wireless literature supplies the method and the motivation.

  • Analysis of Blockage Effects on Urban Cellular Networks models buildings as randomly placed rectangles. We reuse this "Boolean" construction at room scale, with furniture and co-residents as the rectangles.
  • Gupta et al. show that independence errors grow with obstacle size and derive only first-order correlations in 2D. Our setting is small (a handful of sensors in one room), so the joint probability can be computed exactly by ray casting rather than approximated.
  • Joint Outage Probability argues that the joint failure event, not the per-link marginals, is the right quantity. Here the joint event is "patient unobserved."
  • Dependence Control for Reliability Optimization in Vehicular Networks and the "density paradox" in V2V-Assisted V2I MmWave Communication both show that redundancy saturates when failures are shared. We expect the same diminishing return from adding radars.
  • All of these works analyze static snapshots of vehicles or cells, and none asks what the lost observations do to a downstream statistical estimate. That question is ours.

The clinical literature motivates the target.

Passive in-home monitoring would address both gaps, but only if its own missing data are understood.

The sensing literature constrains the design.

  • mRI shows that a single radar gives coarse body tracking (about 125 mm per-joint error versus about 115 mm for RGB). It also shows that wearables are more accurate (40.9 mm in the personalized setting) but drift and depend on placement.
  • mRI is a short, single-radar laboratory dataset with no obstacles, so it cannot by itself say anything about correlated blind spots.

Research Program

We would build three components.

Room-scale blockage model. Static obstacles come from a floor plan, or from a Boolean process when no plan exists. Moving occluders are added on top: the patient's own body and co-residents, modeled as cylinders moving along trajectories. The model outputs, for every location and time, the probability that all $K$ sensors are blocked. Because the static and moving components are kept separate, we can also ask a longitudinal question. Static shadows are learnable from a short calibration window. Transient shadows from co-residents vary across weeks, so a calibration window that is too short misestimates them. This is the limited-follow-up problem appearing inside the sensing system itself.

Correlation-aware observation weighting. The model's predicted probability of being observed, conditional on location and activity context, becomes an inverse-probability weight. Each observation is up-weighted by one over its probability of being seen. These weights feed a mixed-effects model of daily mobility summaries: walking speed, time spent upright, and rooms visited, a home proxy for life-space. The key property is that the weights are defined in shadowed zones where the person is never seen. An empirical approach has nothing to estimate from there, because it lacks the overlap between observed and unobserved conditions that missing-data methods need.

Design calculator. Given a floor plan, a sensor budget, and a target precision on the slope of mobility change, the calculator returns the follow-up length required. It reports this both under independence and under correlated blockage, which makes the planning error visible.

Baselines

The naive reference is complete-case analysis. Unobserved periods are simply dropped, and coverage is planned by assuming every sensor is blocked independently with its measured marginal rate. This is current practice, and it is the baseline the core claim says is biased.

The strongest existing competitor is empirical missing-data correction. A pilot week gives an observed-coverage map. Inverse-probability weighting or multiple imputation then assumes data are missing at random given the last-seen location and time of day. This represents what a careful biostatistician would do with no blockage geometry at all. We would run it on the same recordings, with the same pilot length used to calibrate our model, and the same outcome model. To beat it, our method must reduce bias in the mobility-slope estimate and bring confidence-interval coverage closer to nominal, especially in homes with large furniture or co-residents. Matching its accuracy only in open rooms would count as a tie.

The key ablation keeps our full pipeline but replaces the joint blockage probability with the product of per-sensor marginals. Any gain that survives this ablation is attributable to correlation rather than to geometry in general. A second ablation drops the moving-occluder component, to test whether static furniture alone explains the bias.

For the sensing layer itself, the reference numbers are mRI's: radar-only per-joint error of about 125 mm, and 95.8% mean average precision for fused radar, RGB-D, and wearable action detection mRI. For the coverage-prediction or bias-correction task we found no reported result on Lacuna for correlated-blockage-aware passive monitoring. Our baselines therefore define that number rather than compare against one.

Why not just…?

Why not just assume independence and add a safety margin of extra sensors? Margins work only if redundancy behaves multiplicatively. Two results show it does not when obstacles are large. Gupta et al. show that independence overestimates coverage most when obstacles are large. Wu et al. show outage rising again as shared obstructions blind every candidate relay. Our sensor-count sweep tests directly whether observed hours saturate as radars are added. Saturation would show that a margin chosen under independence is systematically too small.

Why not just use standard missing-data methods on pilot coverage maps? This is the real threat. If patients rarely sit in places no sensor sees, and the moving-occluder component is small, the empirical approach will match ours. Our advantage is limited to two situations: zones never observed during the pilot, and transient co-resident shadows that a one-week pilot under-samples. The decisive check below varies pilot length and household composition to find where, if anywhere, the gap appears. If the gap is negligible, the honest conclusion is that the geometric model is useful only as a pre-deployment design tool, and we would report it that way.

Why not just use wearables, which never lose line of sight? mRI documents drift and sensitivity to attachment position, and multi-month adherence to six worn devices is implausible for frail older adults. In trials like Kwan et al., even scheduled visits lost 14.4% of the intervention arm. Wearables still serve as our reference: worn during short validation windows, they provide the unblocked ground truth against which radar-derived trajectories are scored.

Proposed Experiments

Semi-synthetic study on mRI. mRI's synchronized wearable streams give complete pose trajectories, and its sensor positions are known. We would place virtual furniture, sampled from the room-scale Boolean model, into the recording space. We would also add virtual co-residents replayed from other subjects' trajectories. Ray casting then masks the radar and RGB-D frames that would have been blocked. Because ground truth is complete, we can measure exactly:

  • the error in predicted observed time under independent versus correlated models;
  • the bias in derived mobility summaries;
  • how both errors scale with obstacle size and number of sensors.

Longitudinal extension. Daily-routine location patterns mined with DISCOVER-style smart-home data supply realistic $\pi(x)$. We would simulate months of follow-up with a known, planted decline in mobility, then ask which estimator recovers the planted slope.

Small real deployment. Three or four radars would be installed in a few homes for one to two months, with residents wearing inertial sensors during scheduled windows. This checks whether the predicted joint-blind fraction matches the measured one, against the same complete-case and empirical-correction baselines.

Decisive checks. The proposal fails if any of the following holds:

  • correlated and independent coverage predictions differ by little in realistic furnished rooms;
  • the empirical correction recovers the planted decline as well as our weighting at every pilot length;
  • measured joint-blind time in real homes does not track the model.

It succeeds if independence overstates effective follow-up in furnished or shared homes, and if correlation-aware weighting removes most of the resulting slope bias where empirical correction cannot.