When Learned Dose Predictions Conflict with Pharmacological Knowledge: A Theoretical Framework for Reconciling Artificial Intelligence with Mechanistic Pharmacokinetics Download PDF

Artificial intelligence is increasingly capable of generating patient-specific predictions of drug exposure, response, or dose requirements, while pharmacokinetic and pharmacodynamic models provide mechanistically structured estimates derived from drug disposition, exposure–response relationships, prior population information, and patient covariates. These approaches need not converge. Their disagreement creates a clinically important problem because superior retrospective prediction does not necessarily identify the more credible dose for an individual patient, whereas mechanistic plausibility does not guarantee that a model remains applicable when physiology, treatment context, or population characteristics depart from its assumptions. This theoretical reconciliation article examines how such conflicts should be interpreted when learned predictions, mechanistic PK/PD expectations, and therapeutic drug monitoring provide partially overlapping but non-equivalent information. The central argument is that no modality should receive automatic priority. Disagreement should first be diagnosed according to its likely source, including covariate representation, training-support limitations, model-form uncertainty, population transportability, changing patient biology, target definition, or measurement quality. Reconciliation should then depend on the provenance, clinical relevance, temporal proximity, and validation status of the competing evidence. Therapeutic drug monitoring becomes particularly important when valid patient-linked observations can discriminate among previously plausible predictions, but its evidentiary weight depends on sampling informativeness, analytical validity, and the exposure or response target being estimated. The resulting framework treats concordant, discordant, and unresolved states as provisional clinical states subject to monitoring and revision. Dose selection therefore becomes an evidence-updating process rather than a contest between artificial intelligence and pharmacological mechanism.