Research

A Patient World Model for Early Forecasting of Digital Health Campaign Outcomes: Capabilities and LimitsPreprint

Yunlong Wang

Advanced Analytics, IQVIA, Wayne, PA · arXiv preprint, September 2026

Abstract

Digital direct-to-consumer (DTC) health campaigns are typically measured after they end. While current in-flight forecasts rely on standard classifiers with reusable samples, we replace this approach with a compact patient world model. Our architecture maintains a latent state per patient, learning exposure-conditioned dynamics alongside a weekly conversion hazard to smoothly roll forward into future conversion curves. Evaluated on a US campaign featuring 147,173 patients, the model predicted the final week-52 prescription volume from just week 4 with only a 2.9% error given recorded future exposures, whereas the strongest baseline erred by 13.6–33.1%. Crucially, when the future media plan is unknown, the model can imagine its own future exposures and roll out on them to yield equally accurate forecasts. A Fisher-information analysis confirms the necessity of this design. Without our dense next-exposure supervision, performance on rare outcomes severely degrades, driving the error rate from 2.9% up to 5.2–11.1%, even though common outcomes remain unaffected. Finally, we outline the model's limitations, highlighting how strong selection effects prevent these observational rollouts from being interpreted causally.

What the model is

• A latent state per patient, advanced week by week by a sequence module — a two-layer GRU here, but the slot is swappable, and the paper is explicit that the GRU is not the contribution.
• A survival hazard head that emits a weekly conversion hazard, and an auxiliary transition head that predicts next week's exposure summary from the current state and action.
• A recursive survival rollout that multiplies weekly hazards into cumulative-incidence curves. They are monotone and consistent across every horizon by construction, which is what per-horizon classifiers cannot offer.
• Hidden size 128, about 2×10⁵ parameters, trained on CPU inside the certified environment the patient data cannot leave.

Headline result — forecasting from an early cutoff

Table 1 of the paper. Relative error (%) of the forecast final new-to-brand prescription volume, from a cutoff at week τ through week 52, on the held-out test split with the future exposures known. Coherent is the fraction of per-patient curves that never decrease. A perfectly calibrated forecaster still faces a 1σ sampling floor of 3.7%, 4.1%, 4.5% and 7.0% at these four cutoffs, so the world model's errors sit at or below the floor at every one of them: on this test set there is no headroom left to measure.

The operational result — when nobody knows the future media plan

Every row above was handed the exposures that actually happened. That is the fair comparison, and it is not a deployable one: in flight, no one knows next month's plan. The world model is the only method here that can supply it from within. Its transition head generates each next exposure and the rollout consumes its own output, closed loop, seeing nothing after the cutoff.

Table 2 of the paper. Forecasting the same quantity without access to the future media plan. The closed-loop model, which sees none of the future, lands within 0.5–2.5% at every cutoff — matching the same model handed the true future (0.8–2.9%), with both inside the sampling floor, so the two are statistically indistinguishable. Imagining the plan costs nothing here. A model that imagines its own future is more accurate than a strong survival baseline handed the real one. The capability is specific to the rare outcome: on the ten-times-denser diagnosis outcome the closed-loop rollout degrades to 15–27%, and the paper says so.

What it shows

• Structure, not capacity. Coherence and additive error propagation come from the survival rollout rather than from the network. The same LightGBM, given that rollout, goes from a 2,582% error to 13.6% — still an order of magnitude behind the world model, but the structural fix is worth far more than the model class.
• Dense supervision earns its place when events are rare. A Fisher-information comparison of the two loss terms shows the hazard term's curvature carries a factor h(1−h), small precisely because conversions are rare, while the transition term carries no such factor. It is a scaling argument rather than an exact ratio, and it makes a testable prediction. Removing the transition head — one flag — raises volume error from 0.8–2.9% to 5.2–11.1% on the rare outcome and costs nothing on the common one, which is exactly where the argument said the effect would be.
• The forecast survives not knowing the plan. The imagined rollout is as accurate as the true-future one. Two controls locate why: holding the plan fixed and varying only the training objective still separates the models, and exporting the imagined exposures into the GBM's hazard head makes things worse. The value is in the latent state consuming its own rollout, not in the imagined exposures as a feature.
• Scenario simulation has a hard limit. Within observed dose support the simulator behaves sanely. Switching all future exposure off raises predicted conversion from 0.31 to 0.89 — the highest, where a causal reading demands the lowest. It is selection through censoring: exposure stops once a patient converts, so no future exposure is itself evidence of imminent conversion. The baseline simulator inverts the same way. Exposure-conditioned rollouts must not be read causally.

Cite

@article{wang2026patientworldmodel,
  title   = {A Patient World Model for Early Forecasting of Digital
             Health Campaign Outcomes: Capabilities and Limits},
  author  = {Wang, Yunlong},
  journal = {arXiv preprint},
  year    = {2026}
}