Transition-Based Digital Twin Modelling for Alzheimer's Disease under Sparse Longitudinal Data

Christopher Kipps; Rahman Attar; Sofia Michopoulou; Yilin Zhang; Yinyu Huang

arxiv: 2606.09671 · v1 · pith:XOBCKKLLnew · submitted 2026-06-08 · 💻 cs.LG · cs.AI

Transition-Based Digital Twin Modelling for Alzheimer's Disease under Sparse Longitudinal Data

Yinyu Huang , Yilin Zhang , Sofia Michopoulou , Christopher Kipps , Rahman Attar This is my paper

Pith reviewed 2026-06-27 16:59 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords Alzheimer's diseasedigital twinlongitudinal datatransition modelingsparse datapredictive modelingmachine learningADNI

0 comments

The pith

Transition-based modeling of adjacent visits outperforms sequence models for Alzheimer's progression prediction under sparse data.

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

The paper presents a personalized digital twin framework that combines transition-based and sequence-based strategies to forecast Alzheimer's disease outcomes from multimodal longitudinal observations. It demonstrates that modeling local clinical transitions between adjacent visits delivers higher accuracy in predicting cognitive status and diagnostic categories than full-sequence approaches when applied to the sparse, irregular ADNI dataset. A reader would care because this points to a practical way to build subject-specific predictions and uncertainty estimates even when patient visits are infrequent and incomplete. The work also enables what-if scenario analysis for individualized trajectory exploration.

Core claim

The central claim is that a transition-based digital twin framework, integrating cognitive assessments, clinical variables, and MRI-derived phenotypes from ADNI, predicts cognitive status and diagnostic categories with greater accuracy via local transition modeling of adjacent visits than via sequence-based modeling in sparse longitudinal settings, while also quantifying predictive uncertainty and supporting patient-specific what-if analysis.

What carries the argument

The transition-based digital twin that models clinical state changes between adjacent visits as the primary predictive engine, supplemented by sequence modeling for longer-term dependencies.

If this is right

Local transition modeling supplies a more data-efficient predictive strategy than full sequences for irregular clinical visits.
Sequence models retain value specifically for uncertainty-aware long-range trajectory forecasting.
The dual-branch framework supports interpretable patient-specific scenario analysis for disease monitoring.
Aligning the choice of temporal modeling with the sparse structure of clinical data improves overall robustness.

Where Pith is reading between the lines

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

The same transition-focused design could be tested on other neurodegenerative conditions that produce similarly irregular visit patterns.
Varying the degree of data sparsity in controlled simulations would directly test the claimed efficiency edge.
Adding genetic or fluid biomarker streams might further strengthen the digital twin's subject-specific forecasts.
Clinical trial simulators could use the transition layer to generate realistic patient cohorts under different intervention scenarios.

Load-bearing premise

The leak-free subject-level splits on ADNI data represent typical real-world sparse longitudinal observations without hidden selection biases or preprocessing artifacts that would change the performance comparison.

What would settle it

Re-evaluating the same models on an independent longitudinal Alzheimer's dataset with documented different sparsity patterns or preprocessing choices that yields equal or higher accuracy for the sequence branch would falsify the data-efficiency advantage.

Figures

Figures reproduced from arXiv: 2606.09671 by Christopher Kipps, Rahman Attar, Sofia Michopoulou, Yilin Zhang, Yinyu Huang.

**Figure 1.** Figure 1: Architecture of the hybrid personalised digital twin framework. The MLP branch models adjacent short-term transitions, while the BiLSTM-Attention branch captures longer-range longitudinal dependencies and supports uncertainty-aware forecasting and what-if trajectory analysis. pairs (t → t+6 months). For each pair, the input contains the current visit’s clinical, cognitive, and imaging features, and the ta… view at source ↗

**Figure 2.** Figure 2: MLP branch visualisation on the held-out test set. 3.3 MLP branch test visualisation and calibration [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Top-1 reliability plot for the MLP+mRMR diagnosis classifier on the held-out test set. (a) MMSE residual distribution. (b) Multiclass ROC for DX [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: BiLSTM-Attention branch visualisation on the held-out test set. 3.4 BiLSTM-Attention branch For the BiLSTM-Attention digital twin, residuals remained centred around zero, with most MMSE errors falling within roughly ±3 points. The diagnosis branch achieved macro-AUC 0.928 and macro-F1 0.798, demonstrating meaningful class discrimination even though it remained below the MLP. Its role in the framework is th… view at source ↗

read the original abstract

Alzheimer's disease (AD) progression is highly heterogeneous and is typically observed through sparse and irregular longitudinal data, posing challenges for prediction and personalised monitoring. Existing machine learning approaches have improved AD prediction using multimodal data, yet often focus on static classification or cohort-level risk estimation, providing limited support for subject-specific modelling and uncertainty-aware reasoning. To address these limitations, we present a personalised digital twin framework for AD prediction and scenario-based analysis using multimodal longitudinal data. The proposed approach integrates complementary modelling strategies to capture clinical transitions and temporal dependencies across visits. Using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI), including cognitive assessments, clinical variables, and MRI-derived phenotypes, the framework predicts cognitive status and diagnostic categories while quantifying predictive uncertainty and enabling patient-specific what-if trajectory analysis. Evaluation on leak-free subject-level splits demonstrates strong performance in score forecasting and diagnosis classification. In this sparse and irregular ADNI setting, transition-based modelling of adjacent visits achieved higher predictive accuracy than the sequence-based branch, suggesting that local transition modelling may be more data-efficient. While sequence models remain valuable for uncertainty-aware trajectory forecasting, local transition modelling offers a more data-efficient and robust predictive strategy. These findings highlight the importance of aligning temporal modelling strategies with clinical data structure and suggest that transition-based digital twin formulations may provide a practical and interpretable approach for personalised disease forecasting in neurodegenerative disorders.

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.

Desk Editor's Note private letter to a colleague

The paper shows transition modeling of adjacent visits beats sequence models for accuracy on sparse ADNI data and frames this as a practical digital-twin approach, but the work is an application of known techniques with evaluation limited to internal splits.

read the letter

The key point is that transition-based modeling of adjacent visits outperformed the sequence-based branch on subject-level splits of ADNI data, which the authors take as evidence that local transitions are more data-efficient under sparsity and irregularity.

The paper builds a framework that combines both strategies into a digital-twin setup. It takes multimodal inputs (cognitive scores, clinical variables, MRI phenotypes), produces forecasts for cognitive status and diagnosis, adds uncertainty estimates, and supports patient-specific what-if trajectory analysis. The evaluation uses leak-free subject-level splits, which is the right way to handle longitudinal AD data.

What the work does cleanly is run the head-to-head comparison and draw the practical conclusion that sequence models are still useful for uncertainty-aware forecasting while transitions are more robust for point prediction in this regime. That alignment between modeling choice and data structure is useful to note.

The soft spots are straightforward. The abstract gives no numeric results, baseline numbers, or calibration details, so the size of the advantage is hard to judge from the text alone. All results come from the same ADNI cohort with no external test set, which means the superiority claim is a fitted outcome on that distribution. The paper does not introduce new primitives; it integrates existing transition and sequence ideas.

This is for readers who work on longitudinal modeling for neurodegenerative disease or who need subject-specific forecasts with uncertainty. A methods person in clinical ML would get the most from the comparison and the data-efficiency observation.

I would send it for peer review. The central empirical claim is narrow enough and the setup is standard enough that referees can assess whether the advantage holds up once the numbers and any sensitivity checks are in place.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes a personalised digital twin framework for Alzheimer's disease that combines transition-based modelling of adjacent visits with sequence-based temporal modelling on multimodal sparse longitudinal ADNI data (cognitive scores, clinical variables, MRI phenotypes). It reports that the transition-based branch achieved higher predictive accuracy than the sequence-based branch on leak-free subject-level splits, interprets this as evidence that local transitions are more data-efficient under sparsity, and claims the framework supports uncertainty quantification and patient-specific what-if trajectory analysis.

Significance. If the reported performance advantage is quantified with baselines and holds under external validation, the empirical observation that transition-based modelling can be more data-efficient than sequence modelling for sparse AD longitudinal data would be useful for guiding temporal modelling choices in digital-twin applications for neurodegenerative disease. The manuscript's use of subject-level splits is a positive practice that avoids obvious leakage.

major comments (3)

[Abstract] Abstract: the central claim that 'transition-based modelling of adjacent visits achieved higher predictive accuracy than the sequence-based branch' is presented without any numeric metrics (e.g., MAE, accuracy, AUC), baseline comparisons, confidence intervals, or statistical significance tests, rendering the data-efficiency conclusion impossible to assess from the provided text.
[Evaluation] Evaluation description: all reported results derive from models trained and tested on the same ADNI cohort (subject-level splits only); no external validation cohort or hold-out dataset from a different source is mentioned, so the claimed superiority remains a within-distribution fitted outcome rather than a generalisable finding.
[Framework] Framework description: the abstract highlights 'quantifying predictive uncertainty' as a core capability, yet supplies no details on the uncertainty method (e.g., Bayesian, ensemble, conformal), calibration procedure, or any reliability diagrams/metrics, which is load-bearing for the uncertainty-aware claim.

minor comments (1)

[Abstract] The abstract refers to 'personalised digital twin framework' and 'invented_entities' without a concise one-sentence definition of what constitutes the 'digital twin' versus a standard predictive model.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on the abstract, evaluation setup, and uncertainty quantification. We address each major comment point-by-point below, indicating where revisions will be made to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that 'transition-based modelling of adjacent visits achieved higher predictive accuracy than the sequence-based branch' is presented without any numeric metrics (e.g., MAE, accuracy, AUC), baseline comparisons, confidence intervals, or statistical significance tests, rendering the data-efficiency conclusion impossible to assess from the provided text.

Authors: We agree the abstract should be self-contained with quantitative support. The full manuscript reports these metrics (including MAE, accuracy, AUC with confidence intervals and significance tests) in the results section on subject-level splits. We will revise the abstract to include the key numeric values, baseline comparisons, and a brief note on statistical testing to make the data-efficiency claim directly assessable. revision: yes
Referee: [Evaluation] Evaluation description: all reported results derive from models trained and tested on the same ADNI cohort (subject-level splits only); no external validation cohort or hold-out dataset from a different source is mentioned, so the claimed superiority remains a within-distribution fitted outcome rather than a generalisable finding.

Authors: The evaluation uses only the ADNI cohort with subject-level splits, as described. No external cohort from a different source is available or used in the current study. We will add an explicit limitations paragraph acknowledging that the superiority is demonstrated within the ADNI distribution and discuss the value of future multi-cohort validation while noting that subject-level splits already mitigate leakage. revision: partial
Referee: [Framework] Framework description: the abstract highlights 'quantifying predictive uncertainty' as a core capability, yet supplies no details on the uncertainty method (e.g., Bayesian, ensemble, conformal), calibration procedure, or any reliability diagrams/metrics, which is load-bearing for the uncertainty-aware claim.

Authors: Details on the uncertainty method (ensemble-based with calibration) and associated metrics appear in the methods and results sections. We will revise the abstract to briefly specify the uncertainty quantification approach and reference the calibration evaluation to support the claim. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper reports an empirical result from training and evaluating two modeling branches (transition-based vs. sequence-based) on leak-free subject-level splits of the ADNI cohort. The central claim—that transition modeling showed higher accuracy under sparsity—is a direct outcome of this standard ML experiment rather than any derivation that reduces to its inputs by construction. No self-definitional equations, fitted parameters renamed as predictions, load-bearing self-citations, or ansatz smuggling appear in the abstract or described methodology. The result is presented as an observation on this specific dataset and is therefore self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The framework rests on standard supervised learning assumptions plus the representativeness of ADNI; no new physical entities are postulated.

free parameters (1)

model hyperparameters and transition probabilities
Fitted during training on ADNI to achieve the reported accuracy; central to both branches.

axioms (1)

domain assumption ADNI multimodal longitudinal records are representative of real-world sparse AD progression without systematic selection bias
Invoked when claiming generalizability of the data-efficiency finding.

invented entities (1)

personalised digital twin framework no independent evidence
purpose: Subject-specific trajectory modeling and what-if analysis
Conceptual wrapper around the two modeling branches; no independent falsifiable signature provided.

pith-pipeline@v0.9.1-grok · 5780 in / 1301 out tokens · 21897 ms · 2026-06-27T16:59:43.605240+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 1 canonical work pages

[1]

Alzheimer’s Research & Therapy17(1), 125 (2025)

Amato, L.G., Lassi, M., Vergani, A.A., Carpaneto, J., Mazzeo, S., Moschini, V., Burali, R., Salvestrini, G., Fabbiani, C., Giacomucci, G., et al.: Digital twins and non-invasive recordings enable early diagnosis of alzheimer’s disease. Alzheimer’s Research & Therapy17(1), 125 (2025)

2025
[2]

In: Metaverse Technologies in Healthcare, pp

Balasubramaniam, S., Sumina, S., Kumar, K.S., Prasanth, A.: Machine learning based models for implementing digital twins in healthcare industry. In: Metaverse Technologies in Healthcare, pp. 135–162. Elsevier (2024)

2024
[3]

Journal of Big Data 12(1), 1–28 (2025)

Elgammal, Z., Albrijawi, M.T., Alhajj, R.: Digital twins in healthcare: a review of AI-powered practical applications across health domains. Journal of Big Data 12(1), 1–28 (2025)

2025
[4]

Frontiers in Neuroscience18, 1454856 (2024)

Fekonja, L.S., Schenk, R., Schr"oder, E., Tomasello, R., Tomsic, S., Picht, T.: The digital twin in neuroscience: from theory to tailored therapy. Frontiers in Neuroscience18, 1454856 (2024)

2024
[5]

Scientific Reports9(1), 13622 (2019)

Fisher, C.K., Smith, A.M., Walsh, J.R.: Machine learning for comprehensive fore- casting of alzheimer’s disease progression. Scientific Reports9(1), 13622 (2019)

2019
[6]

Applied Sciences14(22), 10121 (2024) A Transition-Based Digital Twin Framework for AD Prediction 13

Hasan Saif, F., Al-Andoli, M.N., Bejuri, W.M.Y.W.: Explainable AI for alzheimer detection: A review of current methods and applications. Applied Sciences14(22), 10121 (2024) A Transition-Based Digital Twin Framework for AD Prediction 13

2024
[7]

In: International Conference on AI in Healthcare

Karaman, B.K., Sabuncu, M.R.: Assessing the significance of longitudinal data in alzheimer’s disease forecasting. In: International Conference on AI in Healthcare. pp. 3–16. Springer (2024)

2024
[8]

JAMIA Open7(3), ooae087 (2024)

Kumar, S., Yu, S.C., Michelson, A., Kannampallil, T., Payne, P.R.: HiMAL: Multimodal hierarchical multi-task auxiliary learning framework for predicting alzheimer’s disease progression. JAMIA Open7(3), ooae087 (2024)

2024
[9]

arXiv preprint arXiv:2002.03419 (2020)

Marinescu, R.V., Oxtoby, N.P., Young, A.L., Bron, E.E., Toga, A.W., Weiner, M.W., Barkhof, F., Fox, N.C., Eshaghi, A., Toni, T., et al.: The alzheimer’s disease prediction of longitudinal evolution (TADPOLE) challenge: results after 1 year follow-up. arXiv preprint arXiv:2002.03419 (2020)

work page arXiv 2002
[10]

The Lancet Neurology13(8), 788–794 (2014)

Norton, S., Matthews, F.E., Barnes, D.E., Yaffe, K., Brayne, C.: Potential for primary prevention of alzheimer’s disease: an analysis of population-based data. The Lancet Neurology13(8), 788–794 (2014)

2014
[11]

Cognitive Neurodynamics18(3), 775–794 (2024)

Odusami, M., Maskeliunas, R., Damasevicius, R., Misra, S.: Machine learning with multimodal neuroimaging data to classify stages of alzheimer’s disease: a system- atic review and meta-analysis. Cognitive Neurodynamics18(3), 775–794 (2024)

2024
[12]

Artificial Intelligence in Medicine145, 102678 (2023)

Qiang, Y.R., Zhang, S.W., Li, J.N., Li, Y., Zhou, Q.Y.: Diagnosis of alzheimer’s disease by joining dual attention CNN and MLP based on structural MRIs, clinical and genetic data. Artificial Intelligence in Medicine145, 102678 (2023)

2023
[13]

Neurotherapeutics p

Ren, Y., Pieper, A.A., Cheng, F.: Utilization of precision medicine digital twins for drug discovery in alzheimer’s disease. Neurotherapeutics p. e00553 (2025)

2025

[1] [1]

Alzheimer’s Research & Therapy17(1), 125 (2025)

Amato, L.G., Lassi, M., Vergani, A.A., Carpaneto, J., Mazzeo, S., Moschini, V., Burali, R., Salvestrini, G., Fabbiani, C., Giacomucci, G., et al.: Digital twins and non-invasive recordings enable early diagnosis of alzheimer’s disease. Alzheimer’s Research & Therapy17(1), 125 (2025)

2025

[2] [2]

In: Metaverse Technologies in Healthcare, pp

Balasubramaniam, S., Sumina, S., Kumar, K.S., Prasanth, A.: Machine learning based models for implementing digital twins in healthcare industry. In: Metaverse Technologies in Healthcare, pp. 135–162. Elsevier (2024)

2024

[3] [3]

Journal of Big Data 12(1), 1–28 (2025)

Elgammal, Z., Albrijawi, M.T., Alhajj, R.: Digital twins in healthcare: a review of AI-powered practical applications across health domains. Journal of Big Data 12(1), 1–28 (2025)

2025

[4] [4]

Frontiers in Neuroscience18, 1454856 (2024)

Fekonja, L.S., Schenk, R., Schr"oder, E., Tomasello, R., Tomsic, S., Picht, T.: The digital twin in neuroscience: from theory to tailored therapy. Frontiers in Neuroscience18, 1454856 (2024)

2024

[5] [5]

Scientific Reports9(1), 13622 (2019)

Fisher, C.K., Smith, A.M., Walsh, J.R.: Machine learning for comprehensive fore- casting of alzheimer’s disease progression. Scientific Reports9(1), 13622 (2019)

2019

[6] [6]

Applied Sciences14(22), 10121 (2024) A Transition-Based Digital Twin Framework for AD Prediction 13

Hasan Saif, F., Al-Andoli, M.N., Bejuri, W.M.Y.W.: Explainable AI for alzheimer detection: A review of current methods and applications. Applied Sciences14(22), 10121 (2024) A Transition-Based Digital Twin Framework for AD Prediction 13

2024

[7] [7]

In: International Conference on AI in Healthcare

Karaman, B.K., Sabuncu, M.R.: Assessing the significance of longitudinal data in alzheimer’s disease forecasting. In: International Conference on AI in Healthcare. pp. 3–16. Springer (2024)

2024

[8] [8]

JAMIA Open7(3), ooae087 (2024)

Kumar, S., Yu, S.C., Michelson, A., Kannampallil, T., Payne, P.R.: HiMAL: Multimodal hierarchical multi-task auxiliary learning framework for predicting alzheimer’s disease progression. JAMIA Open7(3), ooae087 (2024)

2024

[9] [9]

arXiv preprint arXiv:2002.03419 (2020)

Marinescu, R.V., Oxtoby, N.P., Young, A.L., Bron, E.E., Toga, A.W., Weiner, M.W., Barkhof, F., Fox, N.C., Eshaghi, A., Toni, T., et al.: The alzheimer’s disease prediction of longitudinal evolution (TADPOLE) challenge: results after 1 year follow-up. arXiv preprint arXiv:2002.03419 (2020)

work page arXiv 2002

[10] [10]

The Lancet Neurology13(8), 788–794 (2014)

Norton, S., Matthews, F.E., Barnes, D.E., Yaffe, K., Brayne, C.: Potential for primary prevention of alzheimer’s disease: an analysis of population-based data. The Lancet Neurology13(8), 788–794 (2014)

2014

[11] [11]

Cognitive Neurodynamics18(3), 775–794 (2024)

Odusami, M., Maskeliunas, R., Damasevicius, R., Misra, S.: Machine learning with multimodal neuroimaging data to classify stages of alzheimer’s disease: a system- atic review and meta-analysis. Cognitive Neurodynamics18(3), 775–794 (2024)

2024

[12] [12]

Artificial Intelligence in Medicine145, 102678 (2023)

Qiang, Y.R., Zhang, S.W., Li, J.N., Li, Y., Zhou, Q.Y.: Diagnosis of alzheimer’s disease by joining dual attention CNN and MLP based on structural MRIs, clinical and genetic data. Artificial Intelligence in Medicine145, 102678 (2023)

2023

[13] [13]

Neurotherapeutics p

Ren, Y., Pieper, A.A., Cheng, F.: Utilization of precision medicine digital twins for drug discovery in alzheimer’s disease. Neurotherapeutics p. e00553 (2025)

2025