Precision pharmacotherapy requires decisions about what dose should be administered next, not merely forecasts of what is likely to occur under previously observed treatment patterns. Current patient models can integrate pharmacokinetic, pharmacodynamic, laboratory, physiological, and treatment data, yet their predictions may remain associational when treatment is repeatedly adapted in response to evolving patient states. Delayed efficacy and toxicity, incomplete clinical observation, treatment-confounder feedback, unsupported dose actions, and model uncertainty further limit direct translation from patient-state prediction to dosing decisions. This Original Causal Digital-Twin Architecture Article proposes the Counterfactual Dose Twin as a patient-specific, time-indexed computational branch that shares the observed clinical history but propagates outcomes under one explicitly defined future dose trajectory. A dose-twin ensemble compares permissible trajectories through linked observation, state-estimation, causal-specification, PK/PD, delayed-outcome, uncertainty, and governance layers. Hard safety constraints, action-support checks, calibrated uncertainty, abstention, and escalation are positioned before clinician authorization rather than appended to an optimized dose output. Real-time updating is conceived as a governed process in which new observations revise the estimated patient state while model versions, data provenance, overrides, missingness, and emerging outcomes remain auditable. Validation would require software verification, causal-recovery experiments, temporal and missing-data stress tests, external retrospective evaluation, prospective silent-mode assessment, human-factors testing, and context-specific interventional evidence. The proposed architecture is an original conceptual synthesis rather than a validated dosing model, treatment recommendation, autonomous prescribing system, regulatory determination, or deployment-ready clinical workflow.
IJPRAS
