Budgeted Escalation: Making Adherence Alerts Reach a Human

Electronic pill monitors work, but only partway. In a 4,292-patient cluster-randomized tuberculosis trial across 36 Chinese districts, a pill box that sounds an alarm at dosing time cut the share of patient-months with poor adherence from 29.9% to 17.0%. Adding text messages brought it to 13.9% (Effectiveness of Electronic Reminders to Improve Medication Adherence in Tuberculosis Patients). The same trial reports a quieter failure. The boxes correctly logged which patients were missing doses, yet local health workers routinely did nothing with those flags. The monitoring data existed, but no one acted on it. That unused signal is where the next gain in adherence plausibly sits.

We propose to treat that failure as an allocation problem between a central "hub" and local "spoke" clinics. The hub is a district office that sees every box's log. The spokes are the village clinics whose workers must phone or visit the patient. The hub would stop forwarding every flag. Each week it would send each spoke worker a short list, sized to what that worker can realistically handle, of the patients whose outcome a contact is most likely to change. Some contacts would go to a family member rather than the worker, because in many households the person who can actually get the pill taken is a younger relative. The two-sentence version is this: monitoring already detects non-adherence, but the human follow-up step collapses under too many alerts and no incentive to act. A small, ranked, budget-limited escalation list, possibly routed through a family caregiver, should recover part of the effect that detection alone leaves unused.

Core Claim

A budget-constrained escalation policy, run at the hub on data the monitor box already collects, will raise the fraction of flags that receive a completed follow-up. That in turn will lower poor-adherence patient-months below the 13.9% achieved by monitor-plus-text alone. The key ingredient is ranking by expected benefit of contact, not by risk of non-adherence. We learn that benefit from a small randomized exploration step built into the pilot.

The claim rests on two strands of the file. The TB trial motivates it by showing the escalation gap empirically. A game-theoretic model of hospital alliances (A queuing game theory approach to strategies for supplementing medical alliances with internet hospitals) constrains the design. That paper models tertiary hospitals and community centers as strategic players. It finds that community-level effort falls below the no-platform baseline when the digital system's per-use cost to the spoke is high, and it finds that loosely coupled spokes disengage when the hub keeps too much of the credit. We read an unranked flood of alerts as exactly a high per-use cost. We also treat "who gets credited for a recovered patient" as a design variable, not an afterthought.

Research Program

The work has three stages. Each stage produces something usable even if the next one fails.

Stage one builds and checks the ranking model offline. Using the monitor-arm logs of the ISRCTN46846388 trial, which we would request through the trial's data-sharing route, we predict next-month poor adherence (missing at least 20% of doses, the trial's definition) from daily box-opening patterns. The inputs are streaks of missed doses, drift in opening time, and weekend effects. Offline data cannot tell us who benefits from contact, since escalation was never randomized in that trial. So this stage only establishes whether risk can be predicted well enough to beat a simple threshold at a fixed list size.

Stage two is a simulator of spoke behavior. It follows the queueing-game model's structure, in which worker effort falls as alert volume per worker rises and as the worker's share of credit falls. We would calibrate it where possible to the observed non-escalation in the TB trial. We use the simulator only to choose list sizes and credit rules for the pilot, not as evidence of effect, because its parameters are weakly identified.

Stage three is a prospective cluster-randomized pilot in which every arm receives monitor box plus text messages. Within the ranked arms, each week's selection is

$$S_t = \arg\max_{|S| \le k_j} \sum_{i \in S} \hat{u}_{i,t},$$

where $S_t$ is the set of patients escalated in week $t$, $k_j$ is the weekly contact budget of spoke worker $j$, and $\hat{u}_{i,t}$ is the estimated drop in patient $i$'s probability of a poor-adherence month if contacted. Initially $\hat{u}$ is proxied by predicted risk. A randomly chosen tenth of each week's slots goes to flagged patients drawn at random. That randomization is what lets us estimate true benefit and switch to benefit-ranking mid-trial. The idea follows the online bandit logic of Oralytics, which learned brushing-prompt timing with a pooled Thompson-sampling model under sparse data (A Deployed Online Reinforcement Learning Algorithm In An Oral Health Clinical Trial). Oralytics chose prompts to the patient. We choose which patients a human should contact, under a hard human-capacity constraint.

The caregiver channel is the intergenerational part of the proposal, and we give it a specific, testable role. ADAPTS-HCT modeled adolescent transplant patients and their older care partners as coupled agents. It reported roughly a 3% adherence gain over random message delivery on the 171-dyad Roadmap 2.0 data (Reinforcement Learning on Dyads to Enhance Medication Adherence). That work pushed app messages to both members of an urban, high-resource dyad, and its results rest on imputed rewards. We keep the family link but reverse its direction and function. Often an older patient with a younger, phone-literate relative is the norm in the rural settings described by the ethnobotanical surveys (Ethnobotanical Plants from Dhanaula Area of District Barnala, Punjab (India)). There the relative becomes a recipient of the escalation, a second human who can act when the clinic worker does not. We do not build a multi-agent learner for this. The routing choice is a randomized arm.

Measurement discipline comes from antioxidant chemistry. That field reports panels of assays because each common test captures a different mechanism and they disagree systematically. DPPH, a colored radical that fades when neutralized, underestimates bulky polyphenols because of steric hindrance. FRAP, an iron-reduction test, sees only electron transfer (Structural Elucidation and Antioxidant Activities of Proanthocyanidins from Chinese Bayberry Leaves). Box opening is the adherence equivalent of a cheap, biased single assay: it measures lid movement, not swallowing. We therefore pair it with a second, mechanistically different measure on a random subsample, namely video-observed ingestion scored by an ingestion-verification model in the style of AdCare-VLM (AdCare-VLM). We report agreement between the two. This matters because a policy that makes patients open boxes more without swallowing more would look successful on box data alone. We deliberately do not bring in antioxidant biomarkers as adherence measures. The file contains no evidence they track medication-taking.

Baselines

The naive reference point is the monitor-plus-text arm with the trial's default practice, in which all flags are visible and nothing is structured about follow-up. The published figure for this arm is 13.9% poor-adherence patient-months, from the combined arm of Effectiveness of Electronic Reminders to Improve Medication Adherence in Tuberculosis Patients. It is also the strongest existing method on this question, since the file contains no stronger reported intervention for daily oral regimens. We rerun it concurrently in our own districts rather than borrowing the historical number, because secular trends and site differences could otherwise masquerade as effects. Beating it means a statistically significant reduction in poor-adherence patient-months against the concurrent arm, with box and video measures moving in the same direction.

The strongest simple competitor is a threshold rule with the same list size. Each week, forward to each worker the $k_j$ patients with the most missed doses. It uses the same data, the same budget and the same workers, and it tests whether any learned ranking is needed at all.

The third comparison is incentive-only. All flags are forwarded, but spoke workers receive visible credit and a per-completed-follow-up payment matched in cost to the ranked arm's overhead. This isolates the queueing-game hypothesis that credit, not selection, is what matters.

The key ablation compares risk-ranked, benefit-ranked and random selection at identical budget. It isolates whether ranking by benefit of contact adds anything over ranking by risk. The secondary ablation is caregiver-plus-worker versus worker-only routing under the same ranking. All arms share the same monitoring hardware, the same text schedule and the same model-tuning budget. Tuning uses only stage-one data and the pilot's run-in period.

For the dyadic learning approach, the only reported result is ADAPTS-HCT's ~3% gain over random delivery on Roadmap 2.0. That result was measured on imputed rewards for a different population, so it is not a comparable number for TB escalation.

Why not just…?

Why not just forward every flag and tell staff to follow up? That is effectively what the TB trial did, and staff did not escalate. The queueing-game model gives a mechanism: an uncapped alert stream raises the spoke's per-use cost, which drives effort below baseline (A queuing game theory approach). Our incentive-only arm tests the most generous version of this alternative directly.

Why not just use a missed-dose threshold instead of a learned ranking? This is a real threat, and we expect it may match risk-ranking. Missed-dose counts are probably most of the predictive signal. The proposal's bet is narrower: that benefit ranking beats risk ranking, because the patients most at risk may be the least reachable. If the random-exploration slots show that benefit does not vary across risk strata, we will drop the learned component and recommend the threshold rule with a budget cap. The cap alone would still be a useful, cheap finding.

Why not just pay the spoke workers? It may be sufficient. That would be a clean policy result, and the incentive-only arm exists to detect it. If incentive-only matches the ranked arm, the selection machinery is not worth its engineering cost.

Why not just add more patient-facing messages or a dyadic reinforcement learner? The evidence says messaging to patients saturates. Text messages alone did not significantly beat control in the TB trial. Push notifications did not change behavior in hypertensive patients in a microrandomized trial (Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension). ADAPTS-HCT's modest gain came in simulation. Our lever is the human who acts after the message, which these approaches do not touch.

Proposed Experiments and Decisive Checks

The offline check on trial logs asks whether any model beats the missed-dose threshold in precision at the per-worker list size. We report area under the ROC curve and precision-at-$k$, with the threshold rule as the bar. If nothing beats it, stage three proceeds with threshold selection and the benefit-ranking ablation only.

The pilot randomizes districts to the arms above, with the number of districts set by a power calculation. That calculation uses the between-district variance implied by the published arm-level rates. Endpoints are poor-adherence patient-months as defined in the TB trial, escalation completion rate (the fraction of listed patients actually contacted within the week), and box-versus-video agreement on the verified subsample.

The decisive test is mediation. If completion rates rise in the ranked arms but poor adherence does not fall, then escalation was not the bottleneck and the premise is wrong. If completion rises and adherence falls only on box data but not on video, the policy is teaching lid-opening rather than dose-taking. The proposal succeeds only if completion, box adherence and video-verified ingestion all improve over the concurrent monitor-plus-text arm. The benefit-ranked versus risk-ranked comparison at equal budget then tells us whether the learned part earns its place.