Validating medical digital twins for clinical decision support: beyond predictive accuracy

Area of Science:

• Biomedical Informatics
• Computational Biology
• Clinical Decision Support

Background:

• Medical digital twins are complex systems integrating diverse functions like prediction, simulation, and machine learning.
• Current validation approaches for digital twins are often insufficient when used for clinical decision support, especially for comparing interventions.
• Evaluation must be tailored to the specific intended use of the digital twin.

Purpose of the Study:

• To define appropriate validation requirements for medical digital twins used in clinical decision support.
• To address how to specify validation for systems comparing interventions, treatment timings, dosages, or care strategies.
• To propose a framework for validating intervention-oriented digital twins.

Main Methods:

• Reviewing existing validation traditions in forecast verification, causal inference, uncertainty quantification, and decision theory.
• Proposing the 'scientific-instrument' framing as a validation lens for intervention-oriented digital twins.
• Emphasizing the need for validation metrics that extend beyond predictive accuracy to include uncertainty representation, robustness, and decision-level consequences.

Main Results:

• Retrospective accuracy is insufficient for intervention-oriented digital twins; validation must include uncertainty representation, updating stability, and robustness.
• Intervention-oriented digital twins require validation of action-conditioned questions, including counterfactual consistency and clinically weighted error.
• The level of mechanistic support needed should correspond to the clinical claim being made.

Conclusions:

• Medical digital twins used for clinical decision support require explicit validation statements detailing target population, prediction horizon, supported interventions, uncertainty bounds, and failure conditions.
• The 'scientific-instrument' framing provides a pragmatic approach to defining the scope of digital twin outputs for clinical reasoning.
• Validation must be use-driven, ensuring digital twins reliably support clinical decision-making for interventions.