Identifiability-Guided Assessment of Digital Twins in Alzheimer's Disease Clinical Research and Care

Abstract

Digital twins - personalized, data-driven computational models - are emerging as a powerful paradigm for representing and predicting disease trajectories at the individual level. These models have the potential to support diagnosis, monitor disease evolution, and evaluate therapeutic interventions in virtual settings in the context of clinical trials and patient care. Rigorous model assessment is thus critical for its implementation, but medical data are often sparse, noisy, and vary significantly across individuals, making it challenging to determine whether a digital twin optimized on such data is valid. In such settings, identifiability analysis becomes essential for evaluating whether model parameters can be reliably estimated and interpreted. To address this, we investigate how identifiability can support the clinical application of a computational causal digital twin model for Alzheimers Disease (AD), where data sparsity and variability are particularly pronounced. Our results show that the magnitude and distribution of biomarker data influence the parameter practical identifiability, and that constraints on the model structure and parameters can significantly affect identifiability. We also observe differences in identifiability across diagnostic groups, with several parameters showing significantly different values between individuals with AD, mild cognitive impairment (MCI), and cognitively normal (CN) subjects. Uncertainty quantification for identifiable parameters and their corresponding model trajectories provides visual insight into variability in disease progression and reveals mild trends related to biomarker data spread. This study represents a first step toward incorporating identifiability techniques into clinical digital twin frameworks, using a data-driven, interpretable example based on a previously published AD model.