pith. sign in

arxiv: 2604.27217 · v1 · submitted 2026-04-29 · 💻 cs.AI

Toward Personalized Digital Twins for Cognitive Decline Assessment: A Multimodal, Uncertainty-Aware Framework

Bulent Soykan , Gulsah Hancerliogullari Koksalmis , Hsin-Hsiung Huang , Laura J. Brattain This is my paper

Pith reviewed 2026-05-07 09:50 UTC · model grok-4.3

classification 💻 cs.AI
keywords digital twincognitive declineAlzheimer diseasemultimodal fusionuncertainty quantificationlongitudinal modelingpersonalized medicinestate space models
0
0 comments X

The pith

A multimodal uncertainty-aware framework enables personalized digital twins to model individual cognitive decline trajectories from sparse longitudinal data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Cognitive decline varies widely across people, which makes prognosis, trial design, and treatment planning difficult. The paper proposes the PCD-DT framework to build patient-specific digital twins by combining latent state-space models that capture individual temporal dynamics, fusion of clinical, biomarker, and imaging data, and uncertainty-aware validation with adaptive updating. This matters because accurate personal forecasts from irregular data histories could support better decisions in neurodegenerative conditions. On TADPOLE data the combined cognitive-plus-MRI setup produced the lowest next-visit prediction errors for ADAS13 and ventricle volume, beating a last-observation baseline. The authors also note the potential of conditional generative models to handle rare progression patterns.

Core claim

The PCD-DT framework integrates latent state-space models for individualized temporal dynamics, multimodal fusion for clinical, biomarker, and imaging features, and uncertainty-aware validation and adaptive updating to enable robust modeling of patient-specific disease trajectories from sparse, noisy, and irregular longitudinal data, demonstrated by cohort separation and improved next-visit forecasts on TADPOLE trajectories.

What carries the argument

The PCD-DT framework, which fuses latent state-space models for personal temporal dynamics with multimodal data fusion and uncertainty-aware validation to produce patient-specific trajectory models.

If this is right

  • Individual prognosis and treatment planning improve when models account for personal heterogeneity in cognitive decline.
  • Clinical trial design benefits from better forecasts of patient-specific progression rates.
  • Uncertainty-aware components allow reliable operation even when input data are noisy or incomplete.
  • Conditional generative models can generate synthetic examples of underrepresented progression patterns to improve training.
  • The architecture provides a foundation for clinically deployable uncertainty-aware digital twin systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same modeling structure could transfer to other progressive conditions that produce longitudinal biomarker and imaging records.
  • Evaluating predictions over multi-year horizons rather than single next visits would test the framework's ability to capture extended trajectories.
  • Feeding continuous streams from wearables or electronic records into the adaptive updating step could increase real-world robustness.

Load-bearing premise

The TADPOLE cohort and the next-visit prediction task are representative of the sparse, noisy, irregular longitudinal data the framework is intended to handle in real clinical settings.

What would settle it

A prospective study applying the PCD-DT to an independent cohort with irregular visit intervals and comparing its uncertainty-calibrated long-term predictions against observed patient outcomes.

Figures

Figures reproduced from arXiv: 2604.27217 by Bulent Soykan, Gulsah Hancerliogullari Koksalmis, Hsin-Hsiung Huang, Laura J. Brattain.

Figure 1
Figure 1. Figure 1: Overview of the proposed PCD-DT framework. The architecture links multimodal inputs, temporal state estimation, multimodal fusion, synthetic-data view at source ↗
Figure 2
Figure 2. Figure 2: Average longitudinal trends for key biomarkers in TADPOLE. Plots show mean view at source ↗
Figure 3
Figure 3. Figure 3: Training and validation loss curves for the LSTM full model on the view at source ↗
Figure 5
Figure 5. Figure 5: Ranked structural MRI feature importance for ADAS13 next-visit view at source ↗
Figure 6
Figure 6. Figure 6: Spatial projection of the structural MRI feature importance scores view at source ↗
read the original abstract

Cognitive decline is highly heterogeneous across individuals, which complicates prognosis, trial design, and treatment planning. We present the Personalized Cognitive Decline Assessment Digital Twin (PCD-DT), a multimodal and uncertainty-aware framework for modeling patient-specific disease trajectories from sparse, noisy, and irregular longitudinal data. The framework combines three methodological components: (1) latent state-space models for individualized temporal dynamics, (2) multimodal fusion for clinical, biomarker, and imaging features, and (3) uncertainty-aware validation and adaptive updating for robust digital twin operation. We also outline how conditional generative models can support data augmentation and stress testing for underrepresented progression patterns. As a preliminary feasibility study, we analyze longitudinal TADPOLE trajectories and show clear separation between cognitively normal and Alzheimer's disease cohorts in ADAS13, ventricle volume, and hippocampal volume over five years. We further conduct a multimodal next-visit prediction ablation using an LSTM sequence model on 3{,}003 visit-pair sequences derived from TADPOLE, where the combined cognitive plus MRI configuration achieves the lowest standardized RMSE for both ADAS13 (0.4419) and ventricle volume (0.5842), outperforming a Last Observation Carried Forward baseline. A Bayesian tensor modeling component for high-dimensional imaging fusion is also discussed. These results support the feasibility of the proposed architecture while also highlighting the need for stronger uncertainty calibration and longer-horizon predictive evaluation. The PCD-DT framework provides a principled starting point for personalized in silico modeling in neurodegenerative disease. This work positions PCD-DT as a foundational step toward clinically deployable, uncertainty-aware digital twin systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes the Personalized Cognitive Decline Assessment Digital Twin (PCD-DT) framework, which integrates latent state-space models for individualized temporal dynamics, multimodal fusion of clinical/biomarker/imaging features, and uncertainty-aware validation for modeling sparse, noisy longitudinal cognitive decline trajectories. As a preliminary feasibility study on TADPOLE data, it reports descriptive separation of cognitively normal and Alzheimer's cohorts in ADAS13, ventricle volume, and hippocampal volume over five years, plus an LSTM-based multimodal next-visit prediction ablation on 3,003 visit-pair sequences where the cognitive+MRI configuration yields the lowest standardized RMSE (0.4419 for ADAS13, 0.5842 for ventricle volume) versus a Last Observation Carried Forward baseline. It also discusses Bayesian tensor modeling for imaging fusion and conditional generative models for data augmentation, while noting needs for improved uncertainty calibration and longer-horizon evaluation.

Significance. If the full PCD-DT architecture (state-space models, fusion, and uncertainty components) were implemented and shown to outperform baselines on the claimed tasks, the work could offer a principled approach to personalized in silico modeling of heterogeneous neurodegenerative progression, with potential value for prognosis and trial design. The TADPOLE cohort separation provides concrete descriptive grounding, but the current results rest on a proxy LSTM ablation rather than the novel elements, limiting demonstrated significance to a high-level proposal.

major comments (3)
  1. [Abstract] Abstract (preliminary feasibility study paragraph): The headline performance numbers (standardized RMSE 0.4419 for ADAS13 and 0.5842 for ventricle volume) are produced by a standard LSTM sequence model on 3,003 visit-pair sequences; no quantitative results, ablations, or metrics are reported for the proposed latent state-space models, Bayesian tensor fusion, or uncertainty-aware validation components. This severs the link between the reported numbers and the central PCD-DT framework claim.
  2. [Abstract] Abstract and results description: No train/test split strategy, cross-validation details, error bars, or uncertainty calibration metrics are provided for the next-visit prediction task, despite the framework's emphasis on uncertainty-awareness; the text explicitly flags the need for longer-horizon evaluation, indicating the current evidence is insufficient to support feasibility of the full architecture.
  3. [Abstract] Cohort separation analysis (Abstract): The reported separation in ADAS13, ventricle volume, and hippocampal volume is purely descriptive and does not employ the individualized latent state-space dynamics or multimodal fusion, so it does not test the load-bearing methodological components of PCD-DT.
minor comments (2)
  1. [Abstract] Abstract contains a formatting artifact '3{,}003' that should be rendered as 3,003.
  2. [Abstract] The discussion of conditional generative models for data augmentation and stress testing is outlined at a high level but receives no implementation or evaluation details.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review of our preliminary feasibility study. We address each major comment below, providing clarifications on the scope of the reported results while agreeing to revisions that improve alignment between the abstract and the framework's conceptual nature.

read point-by-point responses

  1. Referee: [Abstract] Abstract (preliminary feasibility study paragraph): The headline performance numbers (standardized RMSE 0.4419 for ADAS13 and 0.5842 for ventricle volume) are produced by a standard LSTM sequence model on 3,003 visit-pair sequences; no quantitative results, ablations, or metrics are reported for the proposed latent state-space models, Bayesian tensor fusion, or uncertainty-aware validation components. This severs the link between the reported numbers and the central PCD-DT framework claim.

    Authors: We agree that the reported next-visit prediction metrics originate from an LSTM-based ablation study rather than direct implementation of the latent state-space models, Bayesian tensor fusion, or uncertainty-aware components. The manuscript presents PCD-DT as a high-level framework proposal, with the LSTM serving as a proxy to illustrate the potential benefits of multimodal inputs on TADPOLE data. No quantitative claims are made for the novel components in the current results. We will revise the abstract to explicitly state that these numbers derive from the proxy ablation and to emphasize that full implementation and evaluation of the PCD-DT architecture constitutes planned future work. revision: yes

  2. Referee: [Abstract] Abstract and results description: No train/test split strategy, cross-validation details, error bars, or uncertainty calibration metrics are provided for the next-visit prediction task, despite the framework's emphasis on uncertainty-awareness; the text explicitly flags the need for longer-horizon evaluation, indicating the current evidence is insufficient to support feasibility of the full architecture.

    Authors: The full manuscript details the data preparation and a chronological train/test split for the 3,003 visit-pair sequences in the methods section to prevent leakage, though this was not summarized in the abstract. Error bars can be incorporated in a revision. We explicitly note in the paper the requirements for improved uncertainty calibration and longer-horizon evaluation, consistent with the preliminary nature of the study. We will update the abstract to briefly reference the split strategy and to reinforce that the current evidence supports only initial feasibility rather than full architecture validation. revision: partial

  3. Referee: [Abstract] Cohort separation analysis (Abstract): The reported separation in ADAS13, ventricle volume, and hippocampal volume is purely descriptive and does not employ the individualized latent state-space dynamics or multimodal fusion, so it does not test the load-bearing methodological components of PCD-DT.

    Authors: The cohort separation analysis is included as a purely descriptive examination of TADPOLE trajectories to illustrate data heterogeneity and progression patterns that motivate the need for personalized modeling. It makes no claim to evaluate or test the latent state-space dynamics or multimodal fusion components. We will revise the abstract to explicitly label this analysis as descriptive and to separate it clearly from both the predictive ablation and the proposed PCD-DT methodological contributions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; framework proposal remains independent of reported proxy results

full rationale

The paper proposes the PCD-DT framework with latent state-space models, multimodal fusion, and uncertainty-aware components, then presents a separate preliminary feasibility study using standard LSTM on 3003 TADPOLE visit-pairs for next-visit prediction and descriptive cohort separation analysis. No equations, self-citations, or derivations are shown that reduce the framework claims to the LSTM outputs or TADPOLE inputs by construction. The ablation is explicitly framed as using LSTM rather than the proposed components, and the results are compared to an external LOCF baseline on the same cohort, preserving independence. This matches the default expectation of no circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The framework rests on standard assumptions about longitudinal medical data and introduces the digital-twin construct without external falsification; no machine-checked proofs or parameter-free derivations are supplied.

free parameters (1)
  • LSTM and state-space model hyperparameters
    Fitted to the 3003 TADPOLE visit-pair sequences; exact values and regularization choices are not reported.
axioms (1)
  • domain assumption Sparse, noisy, irregular longitudinal clinical data can be adequately represented by latent state-space dynamics
    Invoked to justify the first methodological component and the TADPOLE analysis.
invented entities (1)
  • Personalized Cognitive Decline Assessment Digital Twin (PCD-DT) no independent evidence
    purpose: To serve as an in-silico model of patient-specific cognitive decline trajectories
    New named construct introduced to organize the three components; no independent external evidence or falsifiable prediction outside the TADPOLE study is provided.

pith-pipeline@v0.9.0 · 5612 in / 1596 out tokens · 56821 ms · 2026-05-07T09:50:59.939239+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Epidemiology of alzheimer’s disease and other dementias: Rising global burden and forecasted trends,

    S. F. Javaid, C. Giebel, M. A. Khan, and M. J. Hashim, “Epidemiology of alzheimer’s disease and other dementias: Rising global burden and forecasted trends,”F1000Research, vol. 10, p. 425, 2021

    work page 2021

  2. [2]

    Global and regional projections of the economic burden of alzheimer’s disease and related dementias from 2019 to 2050: A value of statistical life approach,

    A. Nandi, N. Counts, S. Chen, B. Seligman, D. Tortorice, D. Vigo, and D. E. Bloom, “Global and regional projections of the economic burden of alzheimer’s disease and related dementias from 2019 to 2050: A value of statistical life approach,”EClinicalMedicine, vol. 51, 2022

    work page 2019

  3. [3]

    N. A. of Engineering, N. A. of Sciences Engineering, and Medicine, Foundational Research Gaps and Future Directions for Digital Twins. Washington, DC: The National Academies Press, 2024. [Online]. Avail- able: https://nap.nationalacademies.org/catalog/26894/foundational- research-gaps-and-future-directions-for-digital-twins

    work page 2024

  4. [4]

    Statistical learning for personalized prediction of alzheimer’s disease progression: A survey of methods, data challenges, and future directions,

    G. H. Koksalmis, B. Soykan, L. J. Brattain, and H.-H. Huang, “Statistical learning for personalized prediction of alzheimer’s disease progression: A survey of methods, data challenges, and future directions,”WIREs Computational Statistics, vol. 17, no. 3, p. e70043, 2025, e70043 EOCS-815.R1. [Online]. Available: https://wires.onlinelibrary.wiley.com/doi/a...

    work page doi:10.1002/wics.70043 2025

  5. [5]

    A comprehensive review on early detection of alzheimer’s disease using various deep learning techniques,

    I. Nagarajan and G. Lakshmi Priya, “A comprehensive review on early detection of alzheimer’s disease using various deep learning techniques,” Frontiers in Computer Science, vol. 6, p. 1404494, 2025

    work page 2025

  6. [6]

    Ai-driven innovations in alzheimer’s disease: Integrating early diagnosis, personalized treatment, and prognostic modelling,

    M. B. Kale, N. L. Wankhede, R. S. Pawar, S. Ballal, R. Kumawat, M. Goswami, M. Khalid, B. G. Taksande, A. B. Upaganlawar, and M. J. Umekar, “Ai-driven innovations in alzheimer’s disease: Integrating early diagnosis, personalized treatment, and prognostic modelling,”Ageing Research Reviews, p. 102497, 2024

    work page 2024

  7. [7]

    Predicting alzheimer’s disease progression using multi-modal deep learning ap- proach,

    G. Lee, K. Nho, B. Kang, K.-A. Sohn, and D. Kim, “Predicting alzheimer’s disease progression using multi-modal deep learning ap- proach,”Scientific reports, vol. 9, no. 1, p. 1952, 2019

    work page 1952

  8. [8]

    Predicting changes in brain metabolism and progression from mild cognitive impairment to dementia using multitask deep learning models and explainable ai,

    F. Garc ´ıa-Guti´errez, L. Hern´andez-Lorenzo, M. N. Cabrera-Mart´ın, J. A. Matias-Guiu, J. L. Ayala, and A. D. N. Initiative, “Predicting changes in brain metabolism and progression from mild cognitive impairment to dementia using multitask deep learning models and explainable ai,” NeuroImage, vol. 297, p. 120695, 2024

    work page 2024

  9. [9]

    Multimodal mixing convolutional neural network and transformer for alzheimer’s disease recognition,

    J. Chen, Y . Wang, A. Zeb, M. Suzauddola, Y . Wen, A. D. N. Initiative et al., “Multimodal mixing convolutional neural network and transformer for alzheimer’s disease recognition,”Expert Systems with Applications, vol. 259, p. 125321, 2025

    work page 2025

  10. [10]

    It: An interpretable transformer model for alzheimer’s disease prediction based on pet/mr images,

    Z. Yao, W. Xie, J. Chen, Y . Zhan, X. Wu, Y . Dai, Y . Pei, Z. Wang, and G. Zhang, “It: An interpretable transformer model for alzheimer’s disease prediction based on pet/mr images,”NeuroImage, vol. 311, p. 121210, 2025

    work page 2025

  11. [11]

    Meta analysis of the diagnostic efficacy of transformer-based multimodal fusion deep learning models in early alzheimer’s disease,

    H. Guo, Z. Yang, G. Zhang, L. Lv, and X. Zhao, “Meta analysis of the diagnostic efficacy of transformer-based multimodal fusion deep learning models in early alzheimer’s disease,”Frontiers in Neurology, vol. 16, p. 1641548, 2025

    work page 2025

  12. [12]

    Predicting alzheimer’s disease progression using deep recurrent neural networks,

    M. Nguyen, T. He, L. An, D. C. Alexander, J. Feng, B. T. Yeo, and A. D. N. Initiative, “Predicting alzheimer’s disease progression using deep recurrent neural networks,”NeuroImage, vol. 222, p. 117203, 2020

    work page 2020

  13. [13]

    A state-space approach for longitudinal outcomes: An application to neuropsychological outcomes,

    A. S. Chua and Y . Tripodis, “A state-space approach for longitudinal outcomes: An application to neuropsychological outcomes,”Statistical Methods in Medical Research, vol. 31, no. 3, pp. 520–533, 2022

    work page 2022

  14. [14]

    Unsupervised multimodal modeling of cogni- tive and brain health trajectories for early dementia prediction,

    M. C. Burkhart, L. Y . Lee, D. Vaghari, A. Q. Toh, E. Chong, C. Chen, P. Tiˇno, and Z. Kourtzi, “Unsupervised multimodal modeling of cogni- tive and brain health trajectories for early dementia prediction,”Scientific Reports, vol. 14, no. 1, p. 10755, 2024

    work page 2024

  15. [15]

    Building a human digital twin (hdtwin) using large language models for cognitive diagno- sis: Algorithm development and validation,

    G. Sprint, M. Schmitter-Edgecombe, and D. Cook, “Building a human digital twin (hdtwin) using large language models for cognitive diagno- sis: Algorithm development and validation,”JMIR Formative Research, vol. 8, p. e63866, 2024

    work page 2024

  16. [16]

    Digital twin for neurology: An introduction to a new frontier in healthcare,

    T. Ashraf, M. A. Chisti, and M. M. Raheem, “Digital twin for neurology: An introduction to a new frontier in healthcare,” in2024 21st Learning and Technology Conference (L&T). IEEE, 2024, pp. 284–289

    work page 2024

  17. [17]

    Digital twins support cross-modal and cross-centric classification of mild cognitive impairment,

    L. G. Amato, R. Minino, M. Lassi, G. Sorrentino, E. Troisi Lopez, V . Moschini, G. Giacomucci, A. Grippo, P. Sorrentino, V . Bessiet al., “Digital twins support cross-modal and cross-centric classification of mild cognitive impairment,”Communications Medicine, vol. 6, no. 1, p. 30, 2026

    work page 2026

  18. [18]

    Using ai-generated digital twins to boost clinical trial efficiency in alzheimer’s disease,

    D. Wang, H. Florian, S.-Y . Lynch, W. Robieson, R. Zhuang, C. Kusiak, J. L. Ross, J. R. Walsh, and O. Graff, “Using ai-generated digital twins to boost clinical trial efficiency in alzheimer’s disease,”Alzheimer’s & Dementia: Translational Research & Clinical Interventions, vol. 11, no. 4, p. e70181, 2025

    work page 2025

  19. [19]

    Deep learning for comprehensive forecasting of alzheimer’s disease progression,

    C. K. Fisher, A. M. Smith, and J. R. Walsh, “Deep learning for comprehensive forecasting of alzheimer’s disease progression,”arXiv preprint arXiv:1807.03876, 2018

    work page arXiv 2018

  20. [20]

    Forecasting progression of mild cognitive impairment (mci) and alzheimer’s disease (ad) with digital twins,

    D. Bertolini, A. D. Loukianov, A. Smith, D. Li-Bland, Y . Pouliot, J. R. Walsh, and C. K. Fisher, “Forecasting progression of mild cognitive impairment (mci) and alzheimer’s disease (ad) with digital twins,” Alzheimer’s & Dementia, vol. 17, p. e054414, 2021

    work page 2021

  21. [21]

    Modeling disease progression in mild cognitive impairment and alzheimer’s disease with digital twins,

    D. Bertolini, A. D. Loukianov, A. M. Smith, D. Li-Bland, Y . Pouliot, J. R. Walsh, and C. K. Fisher, “Modeling disease progression in mild cognitive impairment and alzheimer’s disease with digital twins,”arXiv preprint arXiv:2012.13455, 2020

    work page arXiv 2012