Prediction without intervention: why fifteen years of no-show models have not closed the loop on patient adherence

1 Introduction

Consider a patient at 28 weeks of pregnancy who misses a scheduled prenatal visit. Her chronic hypertension goes unmonitored for another month. The blood pressure trend that would have prompted intervention is never measured. By the time she presents, options have narrowed.

This scenario is neither rare nor unpredictable. Maternal mortality review committees in 36 states determined that more than 80% of pregnancy-related deaths in 2017–2019 were preventable (1). In every period the CDC's maternal mortality review committee dashboard reports, from 2017–2019 through 2022, the contributing factor classes documented most often in preventable pregnancy-related deaths include clinical skill and quality of care, knowledge, lack of continuity of care, and lack of access or financial resources (2); the last two surface in daily practice as appointments not kept. A state-level analysis attributed part of the 1997–2012 rise in maternal mortality (4.9%) to the proportion of births to women attending fewer than ten prenatal visits (3).

The research community's dominant answer to this problem has been to predict it. If we can identify, in advance, which patients will miss which appointments, the reasoning goes, providers can intervene. It is a sensible premise. What it has not produced, I will argue, is a system that intervenes.

2 The prediction plateau

No-show prediction may be the most replicated task in clinical machine learning. A 2025 review of 52 publications from 2010 to 2025 found reported accuracy anywhere from 52% to 99.44% (4), and the top of that range is not believable for a behavioral outcome. The methods vary, logistic regression through gradient boosting and neural networks, but the ingredients rarely do: scheduling data, demographics, clinical history. Neither do the findings. Prior no-shows predict future ones. Long lead times hurt (4, 5). Distance, age, and insurance type turn up again and again, and a few models add weather (6) or socioeconomic variables (4). Refinements keep arriving, interpretable models that explain individual predictions (6), cost-sensitive handling of class imbalance (7), and each one shows, at most, an incremental gain on retrospective data.

Then comes each study's closing paragraph. The model could inform targeted outreach. The findings might support future intervention studies (6). What the intervention would be is sometimes named, but who would deliver it and how anyone would know it worked sits outside the study, in every study. Prediction alone has not been shown to change attendance. Prediction coupled to a targeted intervention has. A rapid systematic review found seven randomized trials and one non-randomized study of model-selected interventions: targeted phone calls (three trials, median relative risk 0.61, moderate certainty), text messages (one trial of 158,669 appointments, 0.91, high certainty), patient navigation (one trial, 0.55), and overbooking in three studies with very low certainty of any effect (8).

Nor is prediction confined to research. One common application is overbooking, which treats the prediction as a capacity-management input and may harm the patient who does arrive (8, 9). Where a patient-directed response exists, it is not standardized (10), and even successful deployments may be unable to quantify their own outreach (11); published evaluations of routine deployments are typically uncontrolled before-and-after comparisons (9, 11); and no study has compared targeted with non-targeted reminders head-to-head, leaving the added value of targeting itself unclear (8). The behavioral signal that outreach generates is used in its thinnest form: whether the patient confirmed, which is among the strongest predictors in the models that include it (12). Learning from what interventions do has been demonstrated only in a 60-patient randomized trial (13), a non-profit maternal program (14), and a one-shot message experiment (15). Table 1 lays this out study by study; it was assembled from the three reviews cited (4, 8, 16) and a targeted search described in its note, is not a systematic review, and bounds the claims of absence in this article.

Table 1. No-show and adherence prediction studies, categorized by the stages of a closed loop.

Studies were identified from the three reviews cited in the text (4, 8, 16), their reference lists, and targeted searches of PubMed and Google Scholar conducted between July and September 2026 for no-show and adherence prediction studies reporting an intervention, a deployment, or a feedback mechanism; the search was not systematic, no screening counts were recorded, and the table should be read as the studies examined rather than as an exhaustive census. Rows are ordered from prediction-only studies, through reminder trials without prediction, prediction-targeted trials, routine deployments of prediction, and communication-sensing studies, to the closest approaches to a complete loop; review articles appear as synthesis rows. “None” indicates the stage was explicitly absent from the study design; “not reported” indicates the study did not evaluate it. Effect sizes are as reported by the cited studies and are not comparable across designs. The Noori et al. entry is a preprint.

aOR, adjusted odds ratio; AUC, area under the receiver operating characteristic curve; CI, confidence interval; EHR, electronic health record; OR, odds ratio; QI, quality improvement; RCT, randomized controlled trial; RMAB, restless multi-armed bandit; RR, relative risk.

The claim here is not that prediction has failed or been ignored. It is that the field has proven the value of connecting prediction to intervention, deployed prediction at scale mainly to manage capacity, sensed only the smallest slice of the signal its own outreach produces, and learned from its interventions only at small scale. No-show rates reported in prediction research commonly fall between 10% and 20% (4). Across four decades of studies worldwide, the average is about 23%, from 13% in Oceania to 43% in Africa, and the median reported rate has declined only a few points per decade (5). Whether discrimination has improved over that period is unknown; the review pools performance across years and does not track it (4). My argument does not depend on a plateau. A tenth gradient boosting model on scheduling data will not change attendance, because the field has been optimizing the wrong stage of the pipeline. Prediction was supposed to be an input. Somewhere along the way it became the product.

3 Why the loop stays open

If predicting and then intervening is a valid premise, why has the connection between the two stayed open for fifteen years? Three structural failures, and two conditions that keep them in place.

3.1 Communication platforms are built as delivery channels, not sensing instruments

Nearly every health system sends automated appointment reminders, and every reminder produces behavioral data: whether the patient responds, how fast, on which channel, and with what stated concern (“I do not have transportation,” “I can't miss work again”). Almost all of it is discarded once the reminder has done its administrative job. The one data stream that directly observes the patient's engagement with her care plan between clinical encounters is treated as plumbing rather than signal. Early machine-learning studies of program messages and participant replies have shown that message characteristics are associated with engagement and premature stopping, but that analysis remains retrospective and disconnected from the risk models that drive care (21).

3.2 Behavioral risk and clinical risk are modeled separately

There has been a lack of recognition that how well an individual adheres to their care plan is a strong predictor of future clinical risk: adherence to treatment regimens roughly triples the odds of a good outcome (22), and missed primary care appointments, adjusted for morbidity burden, carry a dose-related increase in all-cause mortality of up to threefold, and more than eightfold in patients with mental health conditions (23). If a pregnant woman stops attending her scheduled prenatal visits, that is not simply a scheduling problem. She is a higher clinical risk. However, in most implemented systems, these two types of models will never exchange features.

3.3 Interventions are fired and forgotten

A score indicating need may trigger an intervention, but few interventions are implemented as experiments with systematic variability; a recent comprehensive review of the no-show prediction literature names the absence of feedback loops from intervention outcomes as an unresolved gap (4), and the feedback column of Table 1 is empty for all but three studies (13–15). Whether an intervention worked for a given patient is rarely measured, so neither the model nor the intervention logic learns, and the system cannot tell which patients benefit from which intervention.

3.4 Why these silos persist

Two conditions sustain these silos. The first is a lack of trust in how models are developed and how they assign numerical risk to patients. Clinicians’ confidence in these predictions remains fragile without interpretable models based on recognizable clinical logic, and providers tend to reject scores they cannot explain (16). Information overload exacerbates this distrust; when decision support tools generate too many alerts perceived as irrelevant, clinicians start to dismiss them, regardless of their actual importance (24).

The second condition is regulatory, in two senses. Patient messages are identifiable data, and the laws that protect them keep those messages out of modeling environments in many organizations. More consequentially, fusing a behavioral score into a clinical risk model that drives escalation is the step that converts an operational scheduling tool into something likely to meet software-as-a-medical-device definitions under FDA clinical decision support guidance and the EU Medical Device Regulation, with the EU AI Act adding obligations for high-risk health systems; and the continuous adaptation the loop requires raises predetermined change control questions the field is still resolving. Closing the loop therefore requires privacy-preserving design, transparency, and a change-control plan as preconditions.

4 Three design principles for closing the loop

The remedy is not a better classifier. It is an architecture that connects sensing, prediction, intervention, and measurement into a single learning loop (Figure 1), defined by three design principles.

4.1 Treat communication infrastructure as a sensing instrument

Every time a patient answers or ignores a reminder, the interaction leaves metadata behind, and those traces form a continually updating stream of behavioral features: response latency and channel choice index engagement; sentiment across messages tracks distress and fading motivation; natural language processing surfaces barriers that no claims feed contains. The case for sensing is not marginal discrimination; it is that a system which knows why a patient is at risk can match the intervention to the barrier. I call this stream “digital exhaust,” a byproduct of routine communication that becomes a real-time view of adherence between visits. The instrumentation exists in every deployed outreach system; what is missing is the decision to treat its output as clinical signal.

4.2 Fuse behavioral prediction into clinical risk

Because engagement behavior predicts clinical trajectory, behavioral adherence prediction belongs inside the clinical risk model as an input feature, not beside it. A model predicting a missed prenatal visit should do more than trigger a scheduling workflow; its score becomes a dynamic feature in the model stratifying that patient's risk of severe maternal morbidity, so the clinical score responds to the behavioral drift that so often precedes deterioration. There is precedent: adding socioeconomic status to Framingham coronary risk scoring did not change discrimination, but it corrected systematic underestimation of risk for low-SES adults and reclassified 15% of them to a higher treatment threshold (25).

Fusing the two models raises design questions any implementation has to answer. If the adherence score is recomputed as communication data arrive, the clinical score can move between visits, but every behavioral feature must then be built only from data that existed before the moment of prediction, or the clinical model ends up learning from signals that arrived after the outcome it was supposed to anticipate. The result looks excellent in retrospect and useless in deployment. Adding an input also unsettles a clinical model's calibration, and the adherence score is itself an estimate, and its error should propagate into the clinical prediction. Behavioral signals also drift quickly: a deployed no-show model lost discrimination during severe weather (10), and dataset shift in deployed clinical models is well described (26). The fused model needs drift monitoring on both inputs and a retraining cadence set by the faster of the two.

Fusion also creates a problem that drift monitoring does not solve. Once the adherence score both feeds the clinical model and triggers outreach, it stops being an observational predictor and becomes a quantity the system manipulates. The clinical model learned the relationship between adherence and outcome in a world where no one intervened on adherence; in the world the loop creates, high-risk patients are contacted and escalated, and the more effective the intervention, the weaker the association. A naive retraining cycle reads that attenuation as evidence that adherence no longer matters. The remedy is the experimental structure of Section 4.3: when delivery is randomized, assignment enters the fused model as a covariate and the adherence score's contribution to clinical risk stays identifiable. A loop that acts on its own predictions must randomize how it acts.

4.3 Treat every intervention as an experiment

What should a high-risk prediction trigger? A tiered response: automated nudges for simple cases, interactive outreach when the nudge fails, a human coordinator for the barriers automation cannot resolve. The point is that every intervention, at every tier, runs as an experiment, with variation built in and results measured.

Measured against what? “Little e” is platform engagement: the opens, replies, and clicks; “Big E” is the health behavior itself: an appointment kept or a prescription refilled (27). And a risk score answers who is likely to miss an appointment, not which intervention will change that: the first predicts outcome risk, the second predicts treatment effect, which varies across patients and can be learned only from variation in what they receive (28). Systematic variation, as in micro-randomized designs (29), generates the data from which intervention effects can be estimated for different kinds of patients; without it, the loop can identify risk but never learn what reduces it.

4.4 What this architecture adds to existing paradigms

The loop in Figure 1 should look familiar to readers of the learning health system literature. The Institute of Medicine defined a learning health system as one in which knowledge is captured as a by-product of care delivery and fed back into practice (30); the cycle proposed here is that idea applied to patient adherence. What the literature leaves unspecified is the mechanism: what to sense, how to combine signals, what to measure for the individual patient. The reminder trials in Table 1 produced a validated model and an effective intervention, tested separately, with no mechanism for linking them (18).

Three other paradigms each hold one piece. Just-in-time adaptive interventions have worked out when and how to intervene at the individual level, by sensing from the patient's own device (31), but have not taken on clinical risk. Clinical decision support puts a prediction in front of a clinician at the point of care (24), then stops. Digital phenotyping quantifies behavior moment by moment (32), but from the patient's phone rather than the health system's outreach platform. Set against adherence, these four are stages of one loop, and the two joints this article adds are missing from all of them: the outreach channel as a sensing instrument, and behavioral prediction as an input to clinical risk (Supplementary Table 1).

To be explicit about the standing of the claims in this article: it is established that adherence to treatment regimens predicts clinical outcomes and that missed appointments predict mortality (22, 23), that message characteristics are associated with engagement and disengagement (21), that adding a non-clinical input can correct a clinical risk score's calibration for a disadvantaged group (25), and that prediction-targeted interventions can reduce missed appointments in randomized trials (Table 1). This article proposes the connection. What remains to be validated is everything that follows from it: the fusion and sensing gains, transferability across institutions, and whether the closed loop prospectively improves attendance and outcomes (Section 5.2).

5 Discussion

5.1 Objections and hard problems

5.1.1 Surveillance and consent

A system that analyzes the content and metadata of patient messages to predict medical need is a different thing from one that analyzes claims. Patients message their care team in confidence, not to contribute training data, and the adherence-prediction literature recognizes privacy and consent as acute ethical tensions (16). A closed loop therefore needs an explicit consent architecture: purpose limitation and data minimization, tiered consent that lets a patient accept reminders without accepting analysis, an opt-out that does not diminish care, audit logging of secondary use, and structural separation of the messaging environment from the modeling environment.

5.1.2 Equity and the silence problem

Patients without a stable phone, a data plan, stable housing, or English produce little digital exhaust, so a system built on the digitally active knows least about its highest-barrier patients (16). Silence is a signal in its own right, not an empty field: data missing through digital exclusion are missing not at random, and a model trained only on the digitally active will misestimate risk for the rest. The risk is manageable by design: a model checked for calibration in Black and white patient strata before deployment drove live reminder calls in a safety-net system, reducing no-shows predominantly among Black patients and narrowing the disparity (20). Without that discipline, the closed loop amplifies existing inequities with algorithmic confidence (16, 33).

5.1.3 The evidence bar

A loop that converts every elevated score into a coordinator task only moves the bottleneck and reproduces the alert fatigue that already undermines clinical decision support (24); the tiers are load-bearing, reserving human attention for the cases that need it, with the contributing factors attached. Retrospective discrimination, the currency of the prediction literature, will never be sufficient for a system that acts on its own predictions, in a literature where even external validation remains limited (16). Closed-loop systems must be evaluated prospectively, yet a platform that continually reallocates interventions resists a fixed-protocol randomized trial. Micro-randomized trials and the methodology built for just-in-time adaptive interventions are the more promising route (29, 31), and the evaluation framework is itself an open question.

5.2 A research agenda

The architecture breaks down into questions the field can answer now. Fusion: how much does clinical risk prediction improve when a behavioral adherence score is added as an input, and does the gain persist once outreach is delivered on the score? The first half is a retrospective comparison on linked data; the second requires randomized delivery. Sensing: do communication-derived features identify actionable barriers that EHR, claims, and social determinants features cannot, and does matching interventions to those barriers change attendance? Generalizability: do privacy-preserving approaches such as federated learning produce adherence models that hold up at unseen institutions, without which none of this leaves the academic centers where models are built?

A pathway can be stated now, in four phases. First, the retrospective comparisons above, which require only linked data. Second, silent deployment, in which the fused model scores patients prospectively but triggers no action. Third, micro-randomized trials that randomize intervention components at each decision point, which estimate the effect of each tier for each kind of patient and preserve the identifiability of the fused model (29). Fourth, a stepped-wedge rollout across clinics, which measures the effect of the whole loop while every clinic eventually receives it (34).

Return, finally, to the patient at 28 weeks. In the system the field has spent fifteen years building, her missed appointment becomes one more row in next year's training set. In the system it should be building, the unanswered messages three weeks earlier had already moved her adherence score; the transportation barrier she mentioned in passing had already been flagged; the ride had already been offered; and if she still did not arrive, her care team knew the same day, and knew why. The difference between those two systems is not a better algorithm. It is the decision to close the loop.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Claude Opus 4.8 (Anthropic, https://claude.ai) was used for editing assistance and to identify candidate citations during drafting, and for editing assistance and locating passages in source documents during revision. Grammarly (Grammarly Inc.) was used for grammar checking. All references were retrieved and read in full by the author; each claim was checked by the author against the source text; and the author reviewed, edited, and verified all content and takes full responsibility for the final text.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fdgth.2026.1949255/full#supplementary-material

The proposed closed-loop adherence architecture compared with four established paradigms. Supplementary material at the source

Footnotes

Edited by: Md. Mohaimenul Islam, University at Buffalo, United States

Reviewed by: Pankaj Chandre, MIT Art Design and Technology University, India

Hassan Sami Adnan, University of Oxford, United Kingdom