Remote patient monitoring means wearable sensors stream vital signs to clinicians. Current models of when to escalate a patient to intensive monitoring treat each patient alone and assume the radio link has unlimited room. In a dense elderly community or care facility, that assumption breaks. Every escalated patient shares the same airwaves, and multi-hop wireless networks collapse rather than slow down gracefully when overloaded. The pressure on those airwaves also comes from a predictable source. Population aging raises the absolute number of high-risk patients even while each individual's risk falls. Global ischemic heart disease deaths rose 60.43% from 1990 to 2019 while the age-standardized death rate fell 30.8% (Wang et al.). So a facility planned on standardized rates will see more patients cross the escalation threshold at the same time than its network was sized for.
We propose to couple the two layers through a single price on shared airtime. Each patient's escalation decision is solved as in the existing single-patient monitoring models. The difference is that every transmission is charged a price, and that price is set by how saturated the patient's local interference neighborhood is. A demographic decomposition of expected load then tells planners when that price will start to bind. The claim to test has two parts. First, network-blind escalation policies fail abruptly once the elderly population passes a threshold set by local radio capacity, and this happens earlier than age-standardized planning predicts. Second, pricing airtime by the marginal clinical value of each sample moves that failure point further out than either fairness-based or static-priority congestion control.
What exists and what is missing
The monitoring side supplies the patient model. Tiered Service Architecture for Remote Patient Monitoring models a patient as a Markov chain over discretized health states. It derives a threshold policy for switching between an ordinary tier and an intensive tier, and gives a condition under which intensive monitoring is never worth it. Optimal Control for Remote Patient Monitoring with Multidimensional Health States extends this to several physiological dimensions and shows that the switching boundaries skew toward the dimensions that deteriorate fastest. We reuse both formulations as the per-patient subproblem. Both treat monitoring cost as a fixed invasiveness parameter and model each patient in isolation, and that is exactly the assumption we remove.
The network side supplies the constraint and the evidence that it matters. Effects of Interference on Wireless Mesh Networks shows that an unrestrained source can fill a relay with packets the relay then drops, so that roughly half of all flows starve under heavy load. It fixes this by building a conflict graph, in which links are nodes and an edge means two links interfere. It finds maximal cliques, meaning groups of links that cannot transmit at once and so must share 100% of airtime, and allocates rates max-min fairly across the cliques that are bottlenecks. We reuse its conflict-graph cliques as our capacity constraint. We replace max-min fairness, which treats a deteriorating cardiac stream like any other byte stream, with a clinically derived price.
Several results constrain the design and motivate it. Throughput falls with hop count (Analog Network Coding in General SNR Regime), and seven-hop relays have been measured to lose over 80% of effective bandwidth (Bellavista et al.). Dense uncoordinated single-hop transmission drives interference capacities toward zero (Scaling Laws of Single-Hop Cognitive Networks). Together these motivate the problem. Hierarchical Cooperation Achieves Optimal Capacity Scaling in Ad Hoc Networks shows that near-linear capacity scaling is achievable in principle. It requires sensors to quantize and relay one another's signals using detailed channel knowledge, and power-limited wearables cannot do that, so we rule this approach out as a design. Improved Capacity Scaling in Wireless Networks With Infrastructure shows that in dense networks, adding base stations does not improve the capacity scaling exponent. This tempers, though it does not dismiss, the over-provisioning objection discussed below.
On the demographic side, Teoh et al. argue that standardized rates hide how much infrastructure is actually needed. Edentulism in China shows the same divergence: crude prevalence rose about 95% while the age-standardized rate fell (Qin et al.). UKB-MDRMF predicts 1,560 disease codes jointly from UK Biobank data, with median AUC exceeding 0.7, and outperforms single-disease models in 87.5% of conditions. We use it as the source of realistic age-dependent multimorbidity risk, which sets each patient's deterioration dynamics.
Core Claim
Let $N_{a,C}$ be the number of patients in age group $a$ whose sensor links fall in conflict clique $C$. Let $\pi_a$ be the stationary probability that a patient of that age is in the intensive tier under a given policy. Let $r_o$ and $r_i$ be the airtime fractions an ordinary and an intensive stream consume. The expected clique load is then the same weighted sum of age-specific rates and age-group sizes that produces the aging-prevalence paradox:
$$\mathbb{E}[L_C] = \sum_a N_{a,C}\,\big(\pi_a r_i + (1-\pi_a) r_o\big) \le 1 .$$
When $N_{a,C}$ grows faster in older groups than $\pi_a$ falls, the load crosses the clique limit. Past that limit, the conflict-graph analysis predicts collapse rather than gentle degradation. The rise in absolute patient numbers, not the falling individual risk, then determines when the network fails.
Our control replaces the hard coupling across patients with a price $\mu_C$ per unit of airtime in each clique. Each patient $i$ solves its own escalation problem from Tiered Service Architecture for Remote Patient Monitoring. The only change is that the per-step cost of tier $u$ becomes $C_u + \mu_C r_u$, where $C_u$ is the original invasiveness cost. The price is updated from measured utilization $U_C$ with step size $\eta$:
$$\mu_C \leftarrow \max\big(0,\ \mu_C + \eta\,(U_C - 1)\big).$$
This is a standard Lagrangian decomposition of a weakly coupled control problem, and the per-patient problem stays small. Its effect is that a patient escalates only when the clinical value of extra samples exceeds the current congestion price. When the network is uncongested, the price is zero and the policy reduces exactly to the Chandak et al. thresholds. When the network is saturated, airtime goes to the patients whose state would change most with more information, not to whoever has the highest absolute hazard.
Research Program
The work has three parts. The first is a planning analysis. We combine demographic scenarios with age-specific multimorbidity risk to forecast, for a given facility layout, the distribution of how many patients are escalated simultaneously in each clique. We then compare that forecast with the plan an age-standardized analysis would produce.
The second is the priced controller. We will analyze it in two ways. One is to bound its gap to the coupled optimum through the Lagrangian dual, which is a known technique and not our contribution. The other is to characterize how the multidimensional switching surfaces of Optimal Control for Remote Patient Monitoring with Multidimensional Health States shift as $\mu_C$ rises. We limit ourselves to two or three physiological dimensions, where that paper's analysis is tractable.
The third is a cross-layer simulator. Patient Markov chains drive packet generation in a packet-level wireless simulator such as OMNeT++ (used for MAC modeling here). This lets interference, losses and relay drops feed back into what clinicians actually observe. Late detection then raises the probability of reaching the critical state.
The conflict graph models interference as present or absent between link pairs. This is the known weakness of clique methods, because it misses cumulative interference from many distant transmitters. We therefore run the full packet simulator rather than the clique model, and we report how far the clique-predicted collapse point is from the simulated one.
Baselines
The naive reference is the network-blind threshold policy of Tiered Service Architecture for Remote Patient Monitoring, and its multidimensional variant, running over standard contention-based channel access. This is what a deployment built from current monitoring theory would do. No reported numeric result was found on Lacuna for the Chandak et al. policies, since their results are structural theorems. We implement them from the published formulation.
The strongest existing network-side method is model-based clique rate control with max-min fairness from Effects of Interference on Wireless Mesh Networks. That paper reports a 50–100% network throughput gain over uncontrolled routing in a 25-node QualNet simulation, and that roughly half of flows starve without control. It is the strongest competitor because it already prevents collapse; it simply ignores clinical value. We place it under the same Chandak threshold policies so that only the airtime allocation rule differs. We first reproduce a gain in its reported range on a matched 25-node random topology before using it as a baseline.
The second strong baseline is static strict priority. Intensive-tier packets always go before ordinary ones, the way a priority access category in commodity Wi-Fi would handle them, with the same clique rate limits underneath.
The third is the obvious combination: clique allocation using weighted max-min fairness, with weights set to each patient's predicted hazard from a multi-disease survival model trained as in UKB-MDRMF. This baseline doubles as the key ablation. It uses the same risk information as our method but weights by hazard instead of by the marginal value of a sample, so it isolates the price-based escalation rule.
All methods see the same topologies, patient populations, demographic scenarios and random seeds. They receive the same tuning budget, a fixed grid over step size or weights and thresholds selected on held-out seeds. Our method counts as beating a baseline if it lowers critical failures per 1,000 patient-days at a fixed elderly density by a margin outside seed-level confidence intervals. It must also push the density at which failures begin rising at least 20% further out, and do both without higher sensor energy use.
Why not just…?
Why not just add more access points with wired backhaul? In a single care facility this is a real threat, and in steady state it may be the right answer. The theory evidence is mixed. Improved Capacity Scaling in Wireless Networks With Infrastructure shows that dense infrastructure does not change the scaling exponent. A "Small World" Approach to Heterogeneous Networks finds that regular grids of wired infrastructure concentrate traffic at their access points. Neither result tells us the cost in a building. We therefore treat over-provisioning as an experimental arm rather than arguing it away. For each demographic scenario we compute how many access points are needed to keep failures at the uncongested level, with and without pricing. If pricing saves no infrastructure under correlated escalation events, such as a heatwave or a respiratory outbreak, then the planning forecast is the main contribution and the controller is not.
Why not just give intensive-tier traffic strict priority? Strict priority does nothing when most contenders are already intensive, which is exactly the high-density regime the demographic forecast predicts. It also starves ordinary-tier streams. Those streams are the ones that catch early deterioration, so starving them feeds more patients into crisis. If strict priority comes within seed noise of the priced controller across the density sweep, we will report that the simpler mechanism suffices.
Why not just weight the max-min allocation by predicted hazard? This is the obvious intersection idea, and the weighted-clique ablation tests it directly. Hazard measures how likely a patient is to deteriorate, not how much an extra sample changes what clinicians would do. A high-hazard but stable patient gains little from more airtime. We expect the gap to appear mainly among multimorbid patients with uneven dynamics across dimensions, where the switching surfaces of Optimal Control for Remote Patient Monitoring with Multidimensional Health States are most skewed.
Why not just provision from age-standardized rates? Doing so undercounts the patients who drive load, as the 60.43% versus −30.8% ischemic heart disease divergence shows. We include it as the planning baseline and measure how far its predicted collapse density sits from the simulated one.
Proposed Experiments
Patient populations are sampled by age group. Deterioration and recovery probabilities scale with multi-disease hazards estimated on UK Biobank, following UKB-MDRMF. Demographic scenarios hold age-specific risk falling by 10–30% while the population aged 65 and over grows up to threefold, the range described in the research file, and anchor the calibration to the published ischemic heart disease divergence. Topologies cover a 25-node random mesh that matches the Li et al. setup and larger facility and neighborhood layouts with relays of one to seven hops. Metrics are critical failures per 1,000 patient-days, detection delay after a patient's state falls below threshold, clique utilization, fraction of starved streams, and sensor energy.
The first decisive check is whether network-blind policies show an abrupt failure point as elderly density grows, and whether age-standardized planning predicts that point too late. If failures rise smoothly instead, the collapse framing is wrong. The second is whether the priced controller beats hazard-weighted clique allocation and strict priority beyond seed noise at the densities the demographic forecast assigns to the next two decades. The third is whether the infrastructure it saves is material relative to the access-point arm. A positive result on the first check but not the second would still yield a useful, falsifiable planning tool. A negative result on the first would mean demographic growth is not a network problem at the densities studied.
References
- Risk assessment of dietary factors in global pattern of ischemic heart disease mortality and disability-adjusted life years over 30 years
- Tiered Service Architecture for Remote Patient Monitoring
- Optimal Control for Remote Patient Monitoring with Multidimensional Health States
- Effects of Interference on Wireless Mesh Networks: Pathologies and a Preliminary Solution
- Analog network coding in general SNR regime
- Social sharing of connectivity resources: control and encouragement of unselfishness in mobile environments
- Scaling Laws of Single-Hop Cognitive Networks
- Hierarchical Cooperation Achieves Optimal Capacity Scaling in Ad Hoc Networks
- Improved Capacity Scaling in Wireless Networks With Infrastructure
- Global incidence of prostate cancer in developing and developed countries with changing age structures
- Long-term trends in the burden of edentulism in China over three decades: A Joinpoint regression and age-period-cohort analysis based on the global burden of disease study 2019
- UKB-MDRMF: a multi-disease risk and multimorbidity framework based on UK biobank data
- Link 11 Network Simulation Based on OMNeT++
- A "Small World" Approach to Heterogeneous Networks