Three places patient digital twins mislead

Patient digital twins are genuinely useful. They compress a messy medical history into something a model can reason over, they let you simulate counterfactuals, and they occasionally surface subgroups a trial was never powered to find.

They also fail — and they tend to fail with a straight face. Three patterns I keep running into:

1. The twin looks nothing like the patient it's modeling

A twin built from registry data, claims, and a handful of lab values can look plausible on a dashboard and bear no resemblance to the person sitting in front of a clinician. Medication adherence is missing. Social determinants are missing. The last six months of deterioration are missing because the data lags. Any recommendation derived from this twin is a recommendation for a fictional patient who shares a zip code with the real one.

The fix isn't more features. It's being honest in the UI about what the twin is not modeling, so the clinician can weight it accordingly.

2. Distributional twins trained on who made it into the dataset

If your twin population is built from patients who reached a tertiary care center and consented to research, your model has learned the physiology of a specific survivorship-biased cohort. Generalizing to primary care or under-represented populations is not a validation question — it's a first-principles problem that more validation data won't fix.

3. Counterfactuals that smuggle in the treatment effect you were trying to measure

The most seductive failure. You build a twin, you simulate "what if we had given drug B instead of drug A," and the twin confidently predicts an outcome. But the features that drive the prediction are themselves correlated with whoever historically received drug A vs B — prescribers, disease severity, insurance — and the twin has silently encoded the old treatment allocation policy. You end up with a counterfactual that is really a restatement of clinical practice in 2019.

Digital twins that acknowledge their own boundaries — in the data they trained on, in the populations they represent, in the questions they can answer — are the ones clinicians actually use. Everything else gets politely ignored, which is the best-case outcome.