Polygenic Risk Under Success-Triggered Treatment Fatigue

A polygenic risk score is a number that sums thousands of small genetic effects into one estimate of inherited disease risk. For heart disease, such a score finds people whom standard clinical calculators miss. In the UK Biobank, adding it to the guideline calculator lifted discrimination from 0.718 to 0.753 and moved 14.2% of future cases into the high-risk category (Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease). That benefit is computed as if every reclassified person is treated at once and stays treated for life.

Two other literatures say this assumption is wrong, and wrong in a way that falls hardest on the people the score is supposed to help. The pandemic-fatigue literature shows that successful risk reduction lowers perceived risk, and lower perceived risk then erodes adherence (Characterizing Human Collective Behaviors During COVID-19 — Hong Kong SAR, China, 2020). The implementation literature shows that young, low-comorbidity patients receive the least guideline therapy (Implementation of guideline-recommended organ-protective therapy in people with type 2 diabetes and cardiovascular disease, heart failure, or kidney disease: real-world evidence from a German University Hospital). They also show that most decision-support prompts never become signed orders (User Actions within a Clinical Decision Support Alert for the Management of Hypertension in Chronic Kidney Disease). We propose to fit a model of treatment continuation that includes this success-triggered feedback loop to real prescribing records, then recompute what polygenic risk scoring actually buys once realistic uptake and fatigue are included.

Core Claim

Consider people whose risk is mainly genetic: clinically low-risk but polygenically high-risk. A statin quickly brings their modifiable markers, such as cholesterol, down to normal. After that point, the risk shown to clinicians and patients no longer reflects why the person was treated. Our hypothesis is that, among undisclosed-score patients, the hazard of stopping treatment rises as markers improve, independent of how long the person has been on therapy. Because genetically reclassified patients reach "normal" markers sooner and have fewer comorbidities, they should keep a smaller fraction of the published polygenic benefit than patients flagged by the clinical calculator.

If this holds, a headline benefit of 0.093 net reclassification improvement overstates the realized value most in exactly the group it is credited to. That reshapes how polygenic scores should be deployed. It also gives a concrete reason to show the patient's unchanging genetic risk alongside the improving marker, so the displayed risk no longer falls all the way to "normal."

The closest existing work differs on each axis. The cooling-law fatigue model (Quantifying the waning effect of non-pharmaceutical interventions with case study in COVID-19) decays adherence toward zero as a function of time alone, at population scale, for lockdowns. Adherence-aware advice (Learning to Make Adherence-aware Advice) treats the human's probability of following advice as a fixed constant, in synthetic games. The polygenic utility papers measure prediction, not what happens after prediction. Nobody has made adherence depend on the improving risk signal that treatment itself produces, or used that dependence to re-price genetic risk information.

The Model

We generalize the cooling law in one specific way. In the original model, the remaining effect of an intervention, $T_t = We^{-kt}$, cools toward zero, where $W$ is its initial impact and $k$ is a decay rate. Here, a latent perceived-risk state $p_t$ instead relaxes toward the risk currently displayed by the patient's markers, $r_t$. The probability of stopping treatment depends on $p_t$:

$$\frac{dp_t}{dt} = -k\,(p_t - r_t), \qquad \lambda(t) = \lambda_0(t)\,\exp(-\alpha\, p_t + x^\top\beta).$$

In these equations:

  • $k$ is how fast perceived risk tracks displayed risk.
  • $\lambda(t)$ is the instantaneous discontinuation hazard, with $\lambda_0(t)$ a baseline that depends only on time on therapy.
  • $\alpha>0$ is the strength of the feedback.
  • $x$ holds fixed covariates (age, sex, comorbidity count, number of medications), with coefficients $\beta$.

Treatment lowers $r_t$, which lowers $p_t$, which raises $\lambda(t)$. This is the loop documented for lockdowns. Feedback-control accounts of pandemic relaxation (Implicit feedback policies for COVID-19: why "zero-COVID" policies remain elusive) describe the same structure, and so do fear-versus-fatigue contact-rate models (Dynamics of COVID-19 pandemic at constant and time-dependent contact rates).

Unlike a time-only decay, the model predicts restarts after markers rebound. It also makes a counterfactual prediction we can simulate: adding a fixed genetic component to $r_t$ should keep the displayed risk from falling to "normal" and so blunt the feedback.

Research Program

The design rests on one observation that makes it testable with existing data. In current care the polygenic score is not shown to clinicians. Continuation behavior should therefore depend on a patient's clinical profile and marker trajectory, not on their genotype.

That lets us fit the continuation model in a records-rich setting without genotypes and carry it into a genotyped cohort through shared clinical covariates. The fitting setting is the University Hospital Essen FHIR data lake (a hospital electronic health record mapped into a standard interoperability format), whose more than 2 billion records include drug codes and laboratory values (Implementation of guideline-recommended organ-protective therapy...). The genotyped cohort is the UK Biobank.

Stage one extracts statin courses and lipid trajectories from the Essen lake. It also extracts the already-studied cardiorenal therapies (SGLT2 inhibitors and GLP-1 receptor agonists) as a second therapeutic class. We then fit the feedback model and its competitors to these records.

Stage two carries the model into the UK Biobank. There we reproduce the integrated polygenic score and calculator analysis (validation cohort of 272,307, with 7,036 incident cases) and the early-life analysis (Highly elevated polygenic risk scores are better predictors of myocardial infarction risk early in life than later). For each reclassified person we simulate:

  1. initiation, with probability taken from the order-signing rates in the decision-support log study;
  2. continuation, under the fitted dynamics;
  3. events prevented, using standard trial-derived relative risk reduction applied to time actually treated.

If the UK Biobank's linked primary-care prescribing is accessible to us, we will also use it as a direct check on the transported model. The research file does not describe that data source, so we treat it as a bonus rather than a dependency.

Stage three compares, in simulation, three display policies:

  • marker-only risk display;
  • marker-plus-fixed-genetic-risk display (the anchor);
  • budgeted re-prompting in the style of adherence-aware advice, but with non-stationary adherence.

This yields a pre-registered effect size and target subgroup for a later pragmatic trial.

Scope constraints. The polygenic gain is negligible in African-ancestry participants (Polygenic risk score portability for common diseases across genetically diverse populations), so the analysis is restricted to European-ancestry participants and stated as such. Hospital records cannot separate contraindication or patient refusal from unjustified stopping. We therefore report results under two definitions of discontinuation: any gap, and gap without a recorded adverse event or contraindication code.

Baselines

Naive reference: perfect adherence. This is the published polygenic result taken at face value: a C-statistic of 0.753 for the score-enhanced calculator, net reclassification improvement of 0.093, and 14.2% of cases correctly up-classified (Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease). For early heart attack, the comparable figure is AUC rising from 0.70 to 0.79 (Highly elevated polygenic risk scores...). We recompute it on our own UK Biobank extraction so that every later discount is applied to identical people.

Constant-discount baseline. This multiplies benefit by a fixed initiation fraction and a fixed adherence probability. It is the stationary-adherence assumption of Learning to Make Adherence-aware Advice, using 33–38.8% signed initiation orders from User Actions within a Clinical Decision Support Alert.... It is the strongest "simple" competitor for the headline number. Our claim beats it only if the retained-benefit fraction differs between genetically reclassified and clinically flagged patients. A uniform discount cannot produce such a difference.

Time-only decay baseline. This is the cooling-law model with adherence decaying as $e^{-kt}$ regardless of markers. On lockdowns it reached 3.74% mean absolute percentage error versus 26.56% without decay in Germany's first wave, and 3.77% versus 6.27% for the next-best model in North Carolina (Quantifying the waning effect of non-pharmaceutical interventions with case study in COVID-19). No reported result was found on Lacuna for any statin or cardiorenal discontinuation model, so there is no domain-matched number to beat. All models are scored on the same Essen held-out patients.

Flexible statistical baseline. This is the strongest one: a recurrent-event, multi-state survival model (on therapy, off, restarted) with lagged marker values and flexible time effects, tuned with the same cross-validation budget as our model. It contains no structural feedback but can, in principle, learn any marker dependence.

Ablation. The key ablation sets $\alpha=0$ in our model, removing feedback while keeping everything else. A second ablation replaces $r_t$ with "time since last abnormal reading," to test whether improvement itself, rather than mere passage of time since alarm, drives stopping.

Model comparison uses held-out log-likelihood and concordance for discontinuation timing, calibration of predicted continuation by decile, and error on the proportion-on-therapy curve over time. "Beating" the time-only and ablated models means better held-out likelihood and a feedback coefficient whose sign and size replicate across the two therapeutic classes.

Why not just…?

Why not just use the clinical calculator and skip genetic scores? The calculator and the polygenic score are nearly uncorrelated ($r = 0.01$), and the calculator alone misses the up-classified 14.2% of cases. It is also weakest before middle age, when calculator inputs have not yet risen. This alternative could still win our analysis. If implementation-adjusted benefit among reclassified patients collapses toward zero, the honest conclusion is that genotyping is not worth deploying without a change in how risk is displayed. We would report that.

Why not just apply a single realistic discount to the published benefit? A single discount assumes everyone stops at the same rate. The fatigue mechanism predicts that rate depends on how fast markers normalize, and the Essen data already show adherence gradients by comorbidity (32.5% for isolated kidney disease versus 63.9% with three conditions) and by sex (odds ratio 0.76 for women). The differential retained-benefit comparison directly tests whether a uniform discount suffices.

Why not just fit a standard survival model with lagged lab values? This is the real threat, and it may predict discontinuation as well as our structural model. If it does, the feedback mechanism can still be confirmed through its marker coefficient. In that case we would use the flexible model for re-pricing and keep the structural model only for counterfactual display simulations, which the flexible model cannot support without an explicit mechanism.

Why not just run a disclosure trial? A trial is the eventual test, but it takes years and needs a target. This work supplies three things for one: the subgroup where loss is concentrated, the expected size of the anchor effect, and the continuation outcome to power on. Running a trial without these risks repeating the decision-support experience, where most prompts died in the order funnel.

Proposed Experiments and Decisive Checks

Check one: does the feedback exist? In Essen held-out data, the fitted $\alpha$ must be positive and must survive adjustment for time on therapy, for comorbidity, and for adverse-event coding. It must also appear in both statins and cardiorenal agents. The model must beat the $\alpha=0$ ablation on held-out likelihood and on the proportion-on-therapy curve error. If $\alpha$ is indistinguishable from zero, the pandemic-fatigue analogy does not carry over to chronic prevention, and the project reduces to a careful constant-discount re-pricing.

Check two: is the loss concentrated where the score helps? In the UK Biobank, we compute the fraction of perfect-adherence events prevented that survives initiation and fatigue. We compare this fraction across four groups: genetically reclassified, clinically flagged, the early-onset group under 50, and the extreme 98th–99th percentile group. The prediction is a lower retained fraction in the genetically reclassified and early-onset groups than in clinically flagged patients, with bootstrap intervals that exclude equality.

Check three: does a fixed genetic anchor change the simulated picture enough to justify a trial? We compare events prevented per 1,000 screened under three conditions: marker-only display, anchor display, and budgeted re-prompting. We report sensitivity across the plausible range of anchor strength, since that parameter is not identified from current data.

The deliverable is a re-priced estimate of polygenic risk scoring's clinical value under realistic behavior, together with a falsifiable account of where that value leaks away.