Intervention-conditioned human digital twins are designed to predict how a patient would respond to treatments that are not actually administered to that patient. Because these alternatives are never observed, there is no direct way to compare the prediction against reality.
Standard predictive accuracy, which works when outcomes are observable, cannot establish counterfactual credibility in this setting. The authors of the arXiv paper argue that a different kind of validation is needed and propose a 'Concordance' plausibility check.
The abstract does not specify how the Concordance check is implemented or evaluated. It does, however, highlight a core challenge in using digital twins for clinical decisions: the need to assess predictions that, by definition, cannot be verified in the same individual.