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

Author's response

The author team shared

The authors have provided the following additional information in response to the questions raised in this post:

Q1. Clinical decision support: Do you envision identifiability being used prospectively in clinical software to flag low-confidence predictions? From a regulatory and patient safety standpoint, having a built-in mechanism to indicate uncertainty could be valuable for both clinicians and trial sponsors.

• Identifiability analyses have been incorporated, and deemed necessary in some cases, in the modeling pipeline of many disciplines (particularly systems biology). However, there are two main bottlenecks I see in slowing this progress. The first being that tools are not sophisticated or robust enough for larger computational problems, much of this due to poor computational efficiency, making scalability and standardization difficult. Models that operate with problems on a larger scale typically involve high-dimensional parameter spaces and are usually statistical or descriptive rather than mechanistic, rendering identifiability tools impractical. Second, the scarcity of longitudinal data for the validation of disease-related models and the difficulties in constructing models that are clinically interpretable further limit clinical translation. Regardless, uncertainty quantification is vital to report in predictive modeling, and such analyses should be accompanied with an identifiability assessment a priori.

Q2. Biomarker expansion: Have you considered integrating additional biomarker layers, such as proteomic or lipidomic signatures, or electrophysiological measures like EEG to enhance discriminability beyond traditional imaging and cognitive endpoints?

• Alzheimer’s disease models at the cellular scale (Hao et al. 2016) and computational EEG approaches for Alzheimer’s classification (Vicchietti et al. 2023) co-exist, but constructing a longitudinal multimodal model that preserves these mechanistic qualities is not easy for a multitude of reasons. Electrophysiological measures are often riddled with noise with unknown underlying circuit mechanisms, and there remains a gap between known protein accumulation and circuit level cognitive dysfunction in disease modeling. Incorporating mechanisms at both the cellular and whole brain level is similarly difficult as these are operating at completely different timescales, differing by many magnitudes. While multimodal models are always appealing, we would encounter many roadblocks in building a disease model progressing over decades that can realistically include all these features.

Q3. Treatment-response modeling: Could this framework be adapted to evaluate “therapeutic digital twins” that simulate differential drug responses across individuals? Such an extension could be highly relevant for trial stratification or responder enrichment strategies in pharmaceutical development.

• We aimed for this framework to be widely applicable to both disease and therapeutic mechanisms. Our group is currently investigating the treatment responses of several new therapeutic drugs for Alzheimer’s disease, and moving forward, incorporating identifiability-driven uncertainty quantification is certainly within our scope. This work has been accepted for publication in npj Systems Biology and Applications (arXiv:2503.08938).

Q4. Applicability to related neurodegenerative conditions: Given that the approach appears model-agnostic, do you foresee it translating to other heterogeneous disorders such as Parkinson’s disease or vascular dementia, where variability in progression is similarly challenging for prediction models?

• The answer to this question is similarly yes, as we’d expect this type of framework to be applicable to any mechanistic model with a reasonably sized parameter space. Application of the method to Parkinson’s disease would be feasible but I’m unsure as to how effective it would be clinically. The consequence of reduced dopamine to mechanisms at the circuit level and downstream motor functionality is rather well studied, but the manifestation of the disease is incredibly diverse (different levels of control across the body, between subjects, across age) and introduces additional complexity on top of cognitive impairment. Therefore, while the method is applicable if such a model exists, I would be very interested in seeing how uncertainty would propagate to the symptoms level, or whether that would even be interpretable.