After a disaster, drones carrying cellular base stations can restore connectivity. Every placement method we know of decides where to fly from how many users are in an area, or how good their radio links are. A mass-casualty scene breaks that logic. The patients who most need the network are often not the ones whose vital signs already look bad. They are the ones in compensated distress: the body is masking the injury, and only an upstream signal shows it. Examples are a shift in raw ECG waveform shape, or rising cellular-stress fluorescence in tissue. Blood pressure and oxygen saturation look normal until the reversible window has closed.
We propose to plan aerial base station coverage around that window. Each patient carries a small sensor that computes a leading-indicator risk score locally. The drones choose positions and altitudes to maximize the expected number of patients whose silent deterioration is reported while it is still reversible. This differs from maximizing throughput or serving the patients who currently look sickest.
The core claim is that the value of covering a patient depends on three things. The first is how likely it is that something has gone wrong since we last heard from them. The second is how long a warning stays useful before the lagging signs reveal the problem anyway. The third is whether we can still act on it. This yields a simple per-patient weight that rises with time since last contact and then saturates at the indicator's lead time. Plugged into a standard coverage optimizer, it should beat both throughput-driven and acuity-driven deployment on a clinically meaningful metric. It should do so exactly in the regime where lead times are long relative to how often the drones can revisit each patient. The claim also predicts that the advantage vanishes as lead time shrinks to zero, and that prediction can be checked.
What the two directions bring, and the gap
Learned aerial base station deployment is mature on communication metrics. One line forecasts user movement and then adjusts drone positions and power. Trajectory Design and Power Control for Multi-UAV Assisted Wireless Networks forecasts movement from geolocated tweets with an echo state network and uses multi-agent Q-learning, reporting about 17% sum-rate gain over static deployment. HAP-assisted multi-aerial base station deployment via federated deep reinforcement learning trains a federated actor-critic policy on a 6 km by 6 km, 30 GHz urban simulation and reaches 94.6–96.7% of centralized multi-agent capacity.
A second line keeps links alive for a fleet of robots doing some other job. Mobile Wireless Network Infrastructure on Demand repositions relay quadrotors to hold rate margins, such as 1 Mbps at 95% confidence, for a separate search team. A third line models disaster coverage geometrically. Comparing Aerial-RIS- and Aerial-Base-Station-Aided Post-Disaster Cellular Networks models demand around a disaster epicenter and exposes an altitude tension: higher flight widens coverage but causes interference to surviving ground cells.
All three lines motivate our work by showing the machinery exists. All three treat every bit and every user as interchangeable. We reuse their channel models and simulation settings, and we replace their notion of demand.
The diagnostic literature supplies the clinical premise, that leading signals precede lagging ones by usable margins. Real-time monitoring of mitochondrial function in the urethral wall shows cellular distress in swine at 30–40% blood loss, and immediate local signal change at human aortic clamping while systemic vitals stayed deceptive. Deep Learning-Based Approach to Emergency Department Chest Pain Risk Stratification by ECG derives a raw-ECG risk score that is statistically independent of troponin. That score carries an adjusted odds ratio of 3.47 for 30-day mortality, which matters because troponin rises only hours after injury. Novel Plasma Biomarker-Based Model for Predicting Acute Kidney Injury After Cardiac Surgery reports a 48–72 hour lag of creatinine behind kidney injury.
These papers also constrain us. Their results are single-center or tiny in-vivo studies, and none report lead times in a form that transfers to field triage. We therefore treat lead time as a swept parameter rather than a measured constant.
The closest joint work is ARTEMIS. It is a quadruped triage robot that classifies Emergency Severity Index level from standard vital signs, with 74% accuracy on Y-MED. It assumes connectivity, and it relies on exactly the lagging inputs our premise distrusts. Strategic Planning for Setting Up Base Stations in Emergency Medical Systems contributes the optimization method we reuse. It shows that a bounded-response coverage objective is monotone submodular, which means each added base station contributes less than the one before. This property lets a lazy greedy algorithm, which adds one site at a time and skips re-evaluating options that cannot win, guarantee at least $(1-1/e)$ of optimal. That paper works on static ambulance stations, not drones or physiology.
Core Mechanism
We model each patient $i$ as being, at any time, either stable or in an onset of hidden distress. Hidden distress is visible to the leading indicator but not yet to vital signs. The leading score last uploaded from patient $i$ gives a hazard rate $\lambda_i$, the chance per unit time of entering hidden distress. A warning is useful only for a lead time $\Delta_i$. After that, the lagging signs appear, and they will be noticed by rescuers or by the low-rate vitals channel regardless.
Let $\tau_i$ be the time since patient $i$ was last covered. The probability that covering $i$ now catches a distress onset that has not yet been reported and is still reversible is
$$w_i(t) = 1 - \exp\left(-\lambda_i \min(\tau_i, \Delta_i)\right).$$
This assumes a constant hazard between contacts. Onsets older than $\Delta_i$ are excluded because reporting them no longer changes the outcome.
At each planning epoch, the drones choose positions $x$, including altitude, to maximize
$$F(x) = \sum_i w_i(t)\, P_i^{\text{link}}(x),$$
where $P_i^{\text{link}}(x)$ is the probability that patient $i$'s uplink succeeds under the air-to-ground line-of-sight and fading model.
With per-drone coverage sets, $F$ is a weighted coverage function and therefore submodular. We solve it with the lazy greedy method of Ghosh & Varakantham, which keeps the optimizer fixed across all comparisons.
Altitude enters naturally through the coverage sets. A low hover gives one high-weight patient a reliable link. A high hover sweeps many patients whose weights have grown, at the cost of more interference.
Two properties distinguish $w_i$ from acuity weighting. First, a patient who currently scores as stable still gains weight as they go uncovered. Second, that weight stops growing at $\Delta_i$, so the scheduler neither neglects nor over-visits.
Research Program
We will build a simulator, because no dataset in our search file joins drone channels with patient physiology. The radio layer reproduces the urban mmWave setting of the federated deployment paper: 6 km by 6 km, 10–200 m altitude, 30 GHz, line-of-sight and non-line-of-sight path loss. It also models surviving ground cells and interference following the post-disaster stochastic geometry model. Victims are placed with that paper's epicenter-centered inhomogeneous point process.
Competing non-medical traffic from bystanders is drawn from Telecom Italia call and internet records, as used in Dual Attention-Based Federated Learning for Wireless Traffic Prediction. This background load is what makes the "crowd penalty" testable: dense uninjured crowds pull throughput-driven drones away from isolated patients.
Patient trajectories pair a lagging-vitals stream with a hidden onset process. Onset rates follow MIMIC-IV-ED acuity strata. The leading score detects onset $\Delta$ before vitals change, with an adjustable detection quality. Lead time $\Delta$ is swept from minutes, the hemorrhage-like regime, to hours, the cardiac and renal regime. Sensors run inference locally and upload scores plus short waveform snippets. The fleet has energy budgets with recharge, using capsule-airport return costs from Joint Deployment and Coverage Path Planning for Capsule Airports.
The primary metric is the salvage-window detection rate: the fraction of distress onsets reported to the command post before $\Delta$ elapses. Secondary metrics are median detection delay, bystander throughput retained, interference caused to surviving ground cells, and energy used.
Baselines
The do-nothing reference consists of surviving terrestrial cells only, plus drones hovering on a fixed grid. This bounds what any movement buys. The infrastructure-on-demand result of total disconnection for static nodes suggests this floor will be low.
The strongest existing deployment method on its own terms is the federated deep reinforcement learning policy of HAP-assisted multi-aerial base station deployment, with its centralized multi-agent counterpart. On that paper's capacity metric it reports 94.6–96.7% of centralized capacity. We found no reported result on Lacuna for any deployment method on a salvage-window or clinical-detection metric. We will retrain both policies in our simulator on sum-rate reward, using the same episode budget and the paper's aggregation interval of 20 episodes. Beating them means a higher detection rate while retaining a stated share of their bystander throughput, which we will report alongside.
The rate-margin baseline treats every patient sensor as a task robot owed a fixed rate guarantee, in the style of Mox et al.. It is equivalent to uniform weights in our lazy greedy optimizer.
The clinically motivated competitor weights patients by current acuity from vital signs, as an ARTEMIS-style triage classifier would produce.
Two ablations isolate our ingredient. The first weights patients by their current leading-indicator score alone, with no staleness term. The second uses staleness without the $\Delta$ cap, which is age-of-information scheduling scaled by hazard.
Every weighting baseline uses the identical lazy greedy optimizer, channel model, fleet, and energy budget, so only the weights differ. We will also give a learned policy the $w_i$ reward. This checks whether the weights help a reinforcement-learning planner as much as they help greedy.
Why not just…?
Why not just weight users by triage acuity inside an existing optimizer? This is the most natural fix, and in one regime it will match us. When $\Delta$ is near zero, leading and lagging signals coincide, and acuity is the right weight. Our claim is that the two diverge once $\Delta$ exceeds the fleet's revisit interval. In that regime acuity weighting keeps hovering over patients who are already recognized and lets stable-looking ones go stale. The $\Delta$ sweep is designed to show exactly where acuity weighting stops being sufficient. If it never does, the proposal fails.
Why not just maximize throughput or serve the densest users? Throughput objectives reward channel quality and user counts. These favor bystander crowds, so they will underperform on detection unless injured patients happen to coincide with dense traffic. We will measure how often that coincidence occurs under the epicenter-centered victim model.
Why not compute the score on the sensor and send it over any thin link, making placement irrelevant? This is a real threat. We already assume on-sensor inference. If a reliable low-rate channel reached every patient, every scheduler would detect every onset. We will therefore run a "universal thin link" condition with an uplink success probability at every patient's location, and report the success level at which our gains disappear. The proposal matters only where post-disaster coverage holes, the "Swiss cheese" pattern of the ARIS/ABS comparison, make drone presence the bottleneck.
Why not use passive reflecting surfaces instead of active base stations? The same comparison shows that reflecting surfaces need a surviving ground station in line of sight, which is often absent at the epicenter. We include them only as a sensitivity condition.
Decisive Checks
The central experiment plots detection-rate advantage over acuity weighting as a function of the ratio of lead time to revisit interval. The prediction is near-zero advantage below a ratio of about one, and growing advantage above it. That threshold follows from the $\min(\tau_i, \Delta_i)$ cap, since staleness can only matter once it can exceed the time it takes to revisit.
A second check degrades the quality of the leading score. Because $\lambda_i$ comes from that score, a weak indicator should collapse our method toward uniform weighting. The two ablations should show which term carries the gain.
A third check reports the altitude distribution chosen by each scheduler. We expect a measurable shift toward low dedicated hovers for high-weight patients, along with the resulting interference cost to surviving ground cells. This makes the altitude trade-off explicit rather than hidden.
All clinical conclusions are conditional on the simulated lead times. Validating real field lead times requires a prospective study outside this proposal's scope, and we will say so plainly.
References
- Trajectory Design and Power Control for Multi-UAV Assisted Wireless Networks: A Machine Learning Approach
- HAP-assisted multi-aerial base station deployment for capacity enhancement via federated deep reinforcement learning
- Mobile Wireless Network Infrastructure on Demand
- Comparing Aerial-RIS- and Aerial-Base-Station-Aided Post-Disaster Cellular Networks
- Real-time monitoring of mitochondrial function in the urethral wall
- Deep Learning-Based Approach to Emergency Department Chest Pain Risk Stratification by ECG
- Novel Plasma Biomarker-Based Model for Predicting Acute Kidney Injury After Cardiac Surgery: A Case Control Study
- ARTEMIS: AI-driven Robotic Triage Labeling and Emergency Medical Information System
- Strategic Planning for Setting Up Base Stations in Emergency Medical Systems
- Dual Attention-Based Federated Learning for Wireless Traffic Prediction
- Joint Deployment and Coverage Path Planning for Capsule Airports with Multiple Drones