Deadline-Aware Courier Selection for Rural Clinical Referrals

Rural clinics increasingly run small diagnostic models on site and send only a compact summary of findings to a distant hospital or cloud model. One system sends about 6.5 KB per patient instead of several megabytes of raw images. Once the message is that small, the speed of the wireless link no longer matters much. What matters is whether the clinic has any working connection at all before the patient's condition makes the referral useless. In a village with a dead uplink, the useful "relays" are people and vehicles that move: an ambulance, a mobile clinic on its rounds, a community health worker's phone. Each can carry the message until it reaches a place with coverage.

We propose to bring the central idea of opportunistic relay selection into this setting. That idea is to let every candidate forwarder set its own countdown timer, so that the best one answers first and the rest stay silent, rather than coordinating all of them. Here, each nearby device sets its timer from its predicted probability of reaching a connected node before the patient's clinical deadline, not from its signal strength. The message is then handed to that single best carrier. The claim is that this recovers most of the delivery reliability of copying the record onto every passing device, while keeping a patient's data on one or two devices. We evaluate it on realistic rural geography and real screening payloads, and we report results for the worst-served settlements as well as the average.

Core Claim

Classical opportunistic relaying picks the single relay with the best bottleneck hop, the larger of $\min(\text{SNR}_1,\text{SNR}_2)$ across candidates. A simple Cooperative diversity method based on network path selection showed that this single choice achieves the same diversity–multiplexing tradeoff as coordinating all relays. The reason is that the chance all $N$ candidates fail at once falls quickly with $N$. Accurate Performance Analysis of Opportunistic Decode-and-Forward Relaying gave the matching closed-form analysis for decode-and-forward relays. We keep the logic of that result and change what counts as a "good path." For a sparse rural network carrying kilobyte payloads, the relevant failure is not a deep fade during one transmission. It is that no carrier reaches a connected node before the case's deadline.

For a referral of clinical class $c$ with deadline $D_c$, each candidate carrier $r$ in range estimates two things. The first is $q_r$, the probability that the short-range handoff succeeds. The second is $T_r$, its time until it next reaches a node with backhaul, estimated from its route schedule or recent movement history. It then sets its timer as

$$\tau_r = \tau_{\max}\,\bigl(1 - \min\{\,q_r,\ \Pr(T_r \le D_c)\,\}\bigr),$$

where $\tau_{\max}$ is the contention window. The $\min$ mirrors the classical bottleneck rule: a carrier is only as good as the weaker of its handoff and its onward delivery. The first timer to expire takes custody and broadcasts a short flag that silences the others. The custodian repeats the selection at every later contact.

Deadlines come from the urgency fields already present in structured edge-diagnostic outputs. A trauma or obstetric case therefore ranks carriers differently from a routine ophthalmology referral. Because only the deadline-hit probability enters the timer, the protocol knows nothing about clinical content beyond the deadline.

This differs from the closest existing work, Relay-Assisted D2D Transmission for Mobile Health Applications. That paper's RSCF method mixes data importance, trust, and residual energy into a selection score aimed at transmission rate in a generic urban device-to-device simulation. RSCF still optimizes rate over a single hop. We argue rate is the wrong objective once payloads are about 6.5 KB: at the 500 kbps "Rural Low" profile, such a payload takes roughly a tenth of a second. We replace rate with deadline-hit probability over a store-carry-forward path, and we place the protocol in a real healthcare geography with a facility hierarchy.

Research Program

The work has three parts.

Geography. We build the network from the Kebbi State data used in MamaCare AI: A GeoAI Framework for Maternal Healthcare Accessibility Analysis in Underserved Communities. That data covers 1,207 functional GRID3 facilities, WorldPop female population on a 100 m grid, and DHS 2018 indicators, in a state where 62% of women live more than 5 km from a functional clinic. Primary health posts are sources. Secondary and tertiary facilities, plus a sparse set of cellular sites, are backhaul-connected sinks.

Uplink availability at each post comes from a terrain-aware propagation model with on/off outage processes. The file contains no measured rural coverage maps, so we treat this layer as synthetic and sweep its parameters. Carrier mobility also has to be simulated, since the file has no real traces:

Payloads. Messages are the structured JSON outputs of the edge layer in A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings. That system's 100-case cardiac, obstetric, trauma, and ophthalmology benchmark supplies payload sizes and urgency mixes. Its design assumes a single direct uplink, which is exactly the assumption we relax. Per-class deadlines are set by clinician input, and sensitivity bands are reported, because no deadline values are given in the file.

Analysis and secondary setting. We extend the diversity argument to custody. If the $N$ candidates' delivery times were independent, the deadline-miss probability of best-carrier selection would fall like a product of $N$ individual miss probabilities. Correlation is the realistic threat: carriers that share one road share failures. We quantify how quickly diversity collapses as candidates share routes. We also study the analogue of outdated channel information, which here means stale schedule predictions. Impact of partial relay selection on the capacity of communications systems with outdated CSI and adaptive transmission techniques showed that stale information and first-hop-only selection degrade classical selection, and we expect parallel effects. As a secondary hierarchical setting, we map the three tiers of A three-tier AI solution for equitable glaucoma diagnosis across China's hierarchical healthcare system onto source, intermediate, and sink roles. That paper treats inter-tier referral as an offline pathway, and this mapping tests whether the results transfer to a non-emergency, longer-deadline regime.

Baselines

Every method runs on the same geography, mobility realizations, outage processes, payload stream, and random seeds. Each method gets the same budget for tuning its free parameters (timer window, score weights, copy limits) on a held-out set of local government areas. Results are reported on the remaining areas. Beating a baseline means a higher fraction of referrals delivered before their deadline in the bottom decile of settlements, ranked by baseline performance, at an equal or smaller number of data copies per referral. Each comparison uses paired seeds and confidence intervals.

Do nothing. The naive reference is the direct uplink with store-and-retry. The post holds the payload and retries whenever its own link comes up, as the architecture of Ting et al. implies. That paper reports end-to-end latency of 25–38 seconds across its simulated rural profiles. That figure assumes the link eventually exists within seconds, and our outage regimes are designed to test exactly that assumption.

Classical signal-based selection. This is the max-min SNR timer of Bletsas et al. and Tourki et al., applied at each contact. It isolates whether replacing signal strength with deadline-hit probability matters.

Strongest existing method. RSCF is our reimplementation of the importance-, trust-, and energy-weighted selector from the device-to-device m-health paper. No reported result was found on Lacuna for RSCF's numerical performance; the paper shows superiority over MTRS, MSRS, and RRS only in figures. We therefore tune its compromise threshold on our validation areas.

Upper bounds. Epidemic flooding copies the record to every encountered device and bounds achievable delivery from above. A future-knowledge oracle bounds single-custody delivery from above.

Ablations. Three ablations isolate the key ingredient:

  • drop the class deadline and rank carriers by expected time to backhaul;
  • drop the handoff term $q_r$;
  • feed the timer schedule predictions that are deliberately stale.

Why not just…?

Why not just keep retrying the direct uplink? For posts with frequent short outages this may be enough, and we expect it to tie in well-covered areas. The proposal rests on the claim that worst-served settlements have outages longer than clinical deadlines. The do-nothing baseline, broken down by settlement decile, tests this directly. If retry matches our method in the bottom decile across realistic outage parameters, the relay layer is unnecessary and we will say so.

Why not just flood the record to every device? This is the real threat. With 6.5 KB payloads, the bandwidth and energy costs of flooding are small, and flooding will deliver at least as often as any single-custody scheme. Our case against it is exposure: every copy is identifiable patient data sitting on a stranger's phone or a vehicle tablet. The experiment that settles it is the tradeoff between deadline-hit rate and copies. We also add a two-copy random replication baseline, because if blind two-copy spraying matches deadline-aware single custody, the selection logic adds nothing. If flooding's delivery gain in the bottom decile is small relative to its copy count, single custody is justified. If the gain is large, we will report that and turn the method into a selector for a small fixed number of copies.

Why not just use signal-based selection, or RSCF, which already handles medical data importance? Both rank candidates by quantities, signal strength and rate, that barely affect delivery of kilobyte payloads. Neither looks at where a carrier is going. The comparison against these two baselines, together with the deadline ablation, shows whether that matters.

Why not just give every post a satellite terminal? This would remove the problem where it can be funded, and our simulation cannot weigh procurement costs. We include an "upgraded uplink" counterfactual to report how many posts would need upgrading to match the relay scheme's bottom-decile performance. This makes the comparison concrete for planners rather than dismissing the option.

Proposed Experiments

The first experiment sweeps outage severity and carrier density across Kebbi local government areas. For each method we report deadline-hit rate by urgency class, copies per referral, and settlement-level equity: the bottom-decile hit rate and the gap between it and the median. The decisive result is whether deadline-aware single custody closes most of the gap between direct retry and flooding in the bottom decile while staying near one copy per referral.

The second experiment tests diversity collapse. Carriers are progressively forced onto shared roads, and we check whether the gain falls the way correlated-candidate analysis predicts. This tells planners whether to diversify mobile clinic routes, not just add vehicles.

The third experiment injects schedule staleness and first-contact-only information. It measures how gracefully the deadline timer degrades compared with signal-based selection.

Finally, we rerun the best configurations on the three-tier referral mapping with longer deadlines, to test whether the gain disappears when urgency is low. If it does, the method should be targeted only at emergency classes.