Routing Guideline Tasks to the Teammate Most Likely to Finish Them

Hospitals turn clinical guidelines into software prompts that fire at whichever clinician happens to open a patient's chart, and most of those prompts die there. In one primary care deployment, clinicians overrode 59.6% of a mandatory hypertension-in-kidney-disease alert. Only 12–14% of the medication titrations it proposed were ever signed. Not one of the recommended nephrology e-consults was completed (User Actions within a Clinical Decision Support Alert for the Management of Hypertension in Chronic Kidney Disease). Wireless networking faces a structurally similar problem. Phones asked to relay other people's traffic refuse when the request drains their battery. They do cooperate when the requester is socially close. Networks get good throughput by choosing which device relays, based on its remaining capacity and its social tie to the sender, rather than by forcing every device to comply.

This note proposes treating a guideline prompt the same way: as a task to be routed, not an interruption to be imposed. When a patient becomes eligible for a guideline action, the system picks the member of the care team who can most plausibly prepare it. That choice depends on the teammate's current workload and on their working relationship with the responsible clinician. The chosen teammate drafts the order, and the responsible clinician signs it at the visit. The bet is that much of the evidence-practice gap is a delivery problem: the right task reaches a tired person, at the wrong moment, from someone outside their working circle.

Core Claim

The claim is specific and falsifiable. For guideline actions whose preparation can be delegated within scope of practice, routing the preparation step improves the signed-completion rate over three alternatives: the standard interruptive alert, workload-aware throttling of that alert, and fixed-role delegation such as "always send to the pharmacist." The routing uses two inputs: an estimate of each teammate's spare capacity, and the strength of their tie to the responsible clinician, measured from shared-patient history. The routing also keeps per-person burden from concentrating on a few highly connected staff.

The social-tie ingredient is the part that has to earn its place. If tie strength adds nothing beyond role and workload, the proposal reduces to load balancing, and we will say so.

What Prior Work Gives Us and What It Leaves Open

Three strands of work supply the pieces, and none joins them.

The first strand is social-aware wireless allocation, which supplies the method. Exploring Social Networks for Optimized User Association in Wireless Small Cell Networks with Device-to-Device Communications matches users to serving nodes with a distributed swap-matching game. A matching game is one where each side ranks the other, and here one user's choice changes the payoffs of others. That paper reports up to a 63% rate gain over signal-strength-only association. However, it assumes the socially central "important" devices will serve as hubs without ever defecting. A Novel Social Aware Reverse Relay Selection Scheme for Underlaying Multi-Hop D2D Communications does model reluctance: it caps each relay's power by tie strength and lets candidates self-select via timers. Its willingness model, though, is a fixed function of past contact. We reuse both mechanisms, swap matching with burden externalities and self-selection by claiming. We replace static willingness with a willingness estimate learned from behavior that depends on current load.

The second strand, selfish-node research, motivates and constrains the design. Safety Challenges and Solutions in Mobile Social Networks and A Survey on Human-Centric Communications in Non-Cooperative Wireless Relay Networks describe social selfishness: people cooperate inside their circle and defect outside it. The first of these also describes a fairness-efficiency tradeoff: loading central hubs maximizes throughput but exhausts them. The zero completed nephrology e-consults look like out-group defection, which motivates using ties. The fairness-efficiency tradeoff means routing must cap hub load.

Deep reinforcement learning can promote sustainable human behaviour in a common-pool resource problem learned an allocation policy in a behaviorally cloned human sandbox. The policy was stingy when the shared pool was low and generous when it was high, and it sustained 65% of games versus 5% under equal allocation. We borrow two things from it. One is the methodology: clone human behavior, then test policies in simulation before deployment. The other is the idea that how much load hubs absorb should depend on team-wide slack. We deliberately do not borrow its sanctions, because penalizing clinicians is neither ethical nor likely to work.

The third strand, implementation science, supplies the regime and the evidence. Clinical Notes Reveal Physician Fatigue links fatigue to an 18% drop in diagnostic yield per standard deviation. That supports load-dependent willingness. Its note-based detector reaches only 60.1% AUC, and its fatigue scores differ by patient race. Both facts constrain us: note-derived fatigue is too noisy to gate on alone, and any load signal must be audited for equity. Algorithmic Social Intervention shows that network-aware selection of peer leaders doubled HIV outreach. It assumes, however, that selected people comply, which is exactly the assumption we drop.

On timing, the free-text overrides in the hypertension study say clinicians refused to order before seeing the patient. The EMBED trial protocol avoids interruptive pop-ups for the same reason. Our prepare-then-sign split exists to respect that constraint.

Research Program

The work proceeds in three stages, each gated on the previous one.

Stage 1: build a willingness model from EHR audit logs. The logs come from a partner health system and record who opened, pended, signed, or dismissed each guideline-related order. From them we estimate the probability $p_{j}(t)$ that clinician $j$ completes a delegated preparation task $t$. The inputs are:

  • $j$'s role and scope;
  • current load, measured as patients assigned, hours into shift, and open inbox items, with the note-perplexity fatigue score as one weak input;
  • tie strength $w_{ij}$ to the responsible clinician $i$, measured as the number of shared patients and co-signed notes over the prior 90 days.

The gating test in this stage is whether $w_{ij}$ improves held-out prediction of completion beyond role and load.

Stage 2: compare routing policies in simulation. We build a behaviorally cloned simulator of a care team, following the sandbox approach of the common-pool paper, from the same logs. Inside it we compare routing policies. The proposed policy solves a capacitated matching problem. At each decision epoch it assigns eligible tasks to teammates to maximize

$$\sum_{t} p_{a(t)}(t) \quad \text{subject to} \quad \sum_{t:\,a(t)=j} c_t \le \kappa_j(S) \ \text{for every } j,$$

where:

  • $a(t)$ is the teammate assigned task $t$;
  • $c_t$ is the task's estimated minutes of work;
  • $\kappa_j(S)$ is $j$'s load cap, which shrinks for highly central staff when team slack $S$ is low.

We solve this with the swap-matching procedure from the small-cell work. Teammates can also decline or claim tasks, mirroring timer-based self-selection.

Stage 3: a stepped-wedge pilot. A stepped-wedge design rolls the intervention out to clinics in a randomized order over time. The pilot covers one guideline, hypertension titration plus the follow-up metabolic panel in chronic kidney disease, in primary care clinics.

Baselines

The naive reference point is the existing hard-stop alert delivered to the chart opener. The cited deployment reports medication-titration signed completion of 12–14%, initiation completion of 33–38.8%, and a 59.6% override rate (User Actions within a Clinical Decision Support Alert for the Management of Hypertension in Chronic Kidney Disease). Titration completion is the number to beat, re-measured in our own clinics during the control periods of the stepped wedge rather than imported.

The strongest competitor on this specific question is workload-aware throttling. It is the same alert, suppressed or deferred to the next face-to-face visit when the clinician's load estimate is high. It uses the identical load model and timing as our system but never reroutes. It is the strongest competitor because it captures the fatigue and timing explanations of non-adherence without any social machinery. No reported result was found on Lacuna for fatigue-gated alert throttling, so we must measure it ourselves.

The second strong competitor is fixed-role delegation, in which every eligible task goes to a designated pharmacist or nurse. It is strong because it is how team-based care is often run, and it needs no model at all.

The key ablation is our router with tie strength set to zero, which is load-aware matching only. It isolates the social ingredient. A second ablation removes hub load caps, to expose the fairness-efficiency tradeoff directly.

All arms share the same eligible-patient definition, the same pend-then-sign workflow where applicable, and the same load features. In simulation, all arms get equal tuning budgets on a common validation split.

Beating the competitors means a higher signed-completion rate within 30 days of eligibility than both throttling and fixed-role delegation, with confidence intervals excluding zero. It also requires a per-clinician task-load Gini no worse than fixed-role delegation, and no widening of completion gaps across patient race or ethnicity.

Why not just…?

Why not just throttle alerts when the clinician is tired? This is the real threat, and it may capture most of the gain. Throttling addresses the "wrong time" and fatigue failures the hypertension study documents. It cannot address tasks that never reach a willing person: the titration still waits on the same overloaded clinician. The throttling arm and the zero-tie ablation settle the question directly. If our router does not beat throttling, the routing idea fails for this guideline.

Why not just change the default order set? Defaults are cheap and proven for stopping things. Removing creatine kinase from cardiac bundles cut redundant tests by 46.7% (Reducing redundant creatine kinase testing in cardiac injury). But titrating a blood-pressure drug in kidney disease requires reviewing labs and dose history. That is adding work, not removing an option, so no default can complete it. We will still include a "prechecked pended order" default inside the throttling arm, so that defaults are not credited to routing.

Why not just use retrospective quality dashboards like HERACLES? The sepsis system lowered in-hospital mortality from 20.54% to 15.27% without interruptions (An artificial intelligence-powered learning health system to improve sepsis detection and quality of care: a before-and-after study). Dashboards act at the unit level and after the fact, and they leave open who acts on each patient. They are complementary rather than competing. Our pilot uses them as a shared monitoring layer across all arms.

Why not just always delegate to pharmacists? Fixed-role delegation may work well if pharmacist capacity is ample. The comparison will show whether load-dependent and tie-dependent assignment matters once that capacity is contested. If fixed-role delegation matches our router, we will recommend it, because it is simpler.

Decisive Checks

Three results would change the plan. First, if Stage 1 shows tie strength adds no held-out predictive value beyond role and load, the social component is dropped before any pilot. Second, if routing beats throttling in simulation but not in the pilot, the cloned simulator is overfit to logged behavior. That is a useful negative result about offline evaluation of clinical workflow policies. Third, if completion rises but load concentrates on a few central nurses, the hub caps are mis-set. We will report the efficiency-burden frontier rather than a single number.