acceptodds
Under review as a conference paper at ICLR 2027

DeMMO: Cross-Disease Longitudinal Modelling of Digital Mobility Outcomes via Multi-Task Learning

Abstract

Digital mobility outcomes (DMOs) derived from wearable sensors characterise mobility in daily life and offer a promising means of monitoring disease progression. However, existing DMO studies have typically focused on either a single disease or a single visit. To the best of our knowledge, this work is the first to define and study the practical problem of cross-disease longitudinal DMO modelling. We argue that this problem should satisfy at least two requirements. First, the temporal progression of DMOs should be modelled within each disease, as mobility-limiting diseases evolve over time. Second, multiple mobility-limiting diseases should be modelled jointly, as different diseases affect different aspects of human mobility. To address this problem, we propose DeMMO, an interpretable framework for longitudinal, multi-disease, and multi-outcome learning. Its central technical contribution is an interpretable cross-disease and cross-outcome relation-learning mechanism that infers signed relations directly from learned longitudinal DMO–outcome mappings, thereby enabling selective information sharing across cohorts without requiring paired participants. We evaluate DeMMO on the recently released, large-scale, multicentre Mobilise-D dataset, which presents a challenging modelling setting involving longitudinal observations, multiple clinical outcomes, and four participant-disjoint cohorts representing distinct mobility-limiting diseases. Compared with eight strong structural longitudinal and deep-regression baselines, DeMMO achieves the best overall predictive performance and outperforms the baselines for most individual outcomes. Stability selection further identifies reliable longitudinal DMO patterns that can inform subsequent clinical validation and disease monitoring.

Primary area:
interpretability and explainable AI
Keywords:
digital mobility outcomes, longitudinal modelling, multi-task learning, cross-disease learning, interpretable machine learning, wearable sensing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_dXvrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents DeMMO for estimating clinical scores from wearable-derived mobility measures across four diseases and five visits. It combines temporal smoothing, feature selection, and learned relations between disease–outcome models. On Mobilise-D, it reports modest but consistent improvements over eight baselines and uses stability selection to identify reliable mobility features.

Critical strengths
  1. The dataset is great. The study uses a valuable combination of four disease cohorts, repeated follow-up, and common mobility measures. This makes it possible to study which mobility patterns can be shared across diseases, a question that single-disease studies cannot answer.

2)The model is easy to inspect, and the evaluation keeps all visits from each participant in the same split. Its gains across most outcome–visit tasks provide encouraging evidence that the approach is useful.

Critical weaknesses
  1. The methodological advance appears limited. Temporal smoothing and sparse feature selection follow cFSGL, while the parameter reconstruction term is close to AutoTR (Zhou and Yang, KDD 2023). Using complete longitudinal mappings as tasks is a reasonable extension, but the paper does not clearly establish a substantial advance in relation learning.

  2. The experiments do not sufficiently isolate the benefit of cross-disease sharing. The comparison with cFSGL changes both sharing and regularization: even with A fixed to zero, the added relation term penalizes coefficient magnitude. Without controls for this effect and for within-disease sharing, the modest improvement is hard to attribute to the main proposed mechanism. These concerns lead me to lean toward weak reject.

Questions for authors
  1. What technical capability does the relation-learning mechanism add beyond applying AutoTR-style parameter reconstruction to longitudinal disease–outcome mappings?

  2. Under the same tuning protocol, how does DeMMO compare with cFSGL plus an L2 penalty, and with a version that allows sharing only within each disease? These comparisons would help establish the separate benefit of cross-disease sharing.

  3. Does the visit-specific model outperform a model that shares the same coefficients across all visits? This would help show how much the temporal component contributes.

Reviewer_TJjvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DeMMO, a novel interpretable multi-task learning framework for cross-disease longitudinal modelling of digital mobility outcomes (DMOs) derived from wearable sensors. The authors address a significant gap in the literature by proposing the first formal definition and study of cross-disease longitudinal DMO modelling, which requires (1) capturing temporal progression of DMOs within diseases and (2) jointly modelling multiple diseases to uncover shared and distinct mobility patterns. The method is evaluated on the large-scale, multi-centre Mobilise-D dataset, demonstrating superior predictive performance over eight baselines across five clinical outcomes in Parkinson’s, MS, COPD, and PFF. The paper also includes stability selection to identify robust, interpretable DMO patterns

Critical strengths
  • Important and Novel Problem Formulation: The paper claims to be first to formally define and study cross-disease longitudinal DMO modelling, a setting with high clinical relevance given the increasing use of wearables in chronic disease monitoring. This conceptual framing is a genuine contribution.
  • Interpretable Model Design: The use of structured sparsity (sparse group Lasso) and temporal smoothness (fused Lasso) ensures that the model remains interpretable, which is crucial for clinical adoption. The coefficient trajectories provide direct insight into how DMO-outcome relationships evolve over time.
  • Use of a High-Quality, Real-World Dataset: The evaluation on the Mobilise-D dataset which is a large-scale, multi-centre, harmonised resource with real-world mobility data across four diseases, is a major strength. The dataset’s scale and longitudinal depth make it ideal for the proposed task.
  • Comprehensive Baseline Comparison: The inclusion of both classical longitudinal models (cFSGL, FRoTS) and recent deep learning methods (MLP, Rank-N, ACCon) demonstrates a thorough understanding of the landscape and strengthens the empirical foundation.
Critical weaknesses
  • The evaluation compares DeMMO against independently trained models (e.g., cFSGL) and deep learning approaches, but fails to include any baseline that explicitly shares information across diseases—such as a multi-task model with shared latent representations or a joint prior over coefficients. This omission makes it impossible to determine whether performance gains stem from the specific architecture of the learned relation matrix or merely from the general benefit of pooling signals across diseases. A more rigorous ablation is needed to isolate the contribution of the proposed mechanism.
  • The method is evaluated on only five prediction tasks (M=5), leading to a relation matrix with just 10 free parameters. While the paper suggests broad applicability across diseases, it offers no discussion of how the approach would perform with a larger number of outcomes—where the O(M2)O(M2) complexity could lead to overfitting or computational bottlenecks. Without addressing scalability, the generalisability of the framework is overstated.
Questions for authors
  • Why was there no inclusion of a multi-disease baseline that incorporates parameter sharing across diseases—such as a shared temporal structure with disease-specific offsets? Without such a baseline, how can the improvement attributed to the learned relation matrix be disentangled from the general benefits of cross-disease information sharing?
  • Have you examined whether the sign of the learned relations—such as the negative association between PD MDS–UPDRS and COPD FEV1—reverses or changes when clinical outcomes are uniformly rescaled to share the same directionality (e.g., all coded so that higher values indicate greater severity)? If this analysis has been conducted, what did it reveal?
  • How would the computational efficiency and statistical reliability of DeMMO be affected as the number of diseases or outcomes grows—say, to 10 or more? Given that the symmetric relation matrix has O(M2)O(M2) parameters, is this formulation inherently scalable, or would it necessitate additional regularisation, sparsification, or structural constraints to remain effective in larger-scale settings?
  • The model uses only DMOs as predictors, but clinical outcomes are influenced by age, sex, comorbidities, medication use, and device-wearing behaviour—all of which vary across diseases and over time. By omitting these covariates, the model risks attributing changes in outcomes to DMOs that are actually driven by unmeasured confounders. This undermines both predictive validity and interpretability.
Reviewer_pD7NOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DeMMO jointly models five clinical outcomes in four participant-disjoint Mobilise-D cohorts over five visits. Each outcome gets visit-specific linear maps from 24 mobility measures under cFSGL's penalties, and the five coefficient matrices are coupled by a learned symmetric, zero-diagonal relation matrix A through λR/2‖U − UA‖². Over five participant-level splits (seeds 42-46), the paper reports the best mean nMSE and weighted correlation against eight baselines (overall nMSE 0.769 vs. 0.787 for cFSGL), with paired t-tests, per-visit results, non-convex variants and an interpretation of A's signs.

Critical strengths
  1. The formulation is fully specified, and the released optimizer implements it: the relation update matches Eq. 19, and both model classes run unmodified on synthetic data. The stated protocol (participant-level splits, training-only standardization; p.6 l.272-279, l.304-307) guards against leakage. This is why I read the problems below as questions of attribution and inference.

  2. The paper reports per-visit results for all baselines (Tables 5-9) and a negative result for non-convex variants (Table 4). These make the comparison useful to others working with these cohorts, whatever one concludes about the method.

Critical weaknesses
  1. The relation term, presented as the "central technical contribution" (p.1 l.022-026), is AutoTR's self-expressive regularizer (Zhou & Yang, 2023, Eq. 2: λ1‖W − WR‖²_F, r_ii = 0), which this paper's own baseline MAGPP (Zhou et al., 2024) also uses. DeMMO moves it from visits to disease-outcome objectives, adds symmetry and swaps its ℓ1 relation penalty for ℓ2. Sec. 2.3 cites neither paper, and App. A.3 (p.16 l.838) mentions relation learning among time points without saying it is the same term. Earlier relation learners with per-task data (MTRL, UAI 2010; AMTL, ICML 2016) never needed paired participants. The paper's claim to methodological novelty rests on this term, so the case for acceptance has to come from the experiments.

  2. The experiments do not show that relation learning causes the gains, which are small and statistically fragile. The cFSGL comparison does not isolate relation learning, contrary to p.7 l.368. At A = 0, the initialization (p.19 l.1016-1017), the relation term is exactly a ridge penalty λR/2‖U‖². Only DeMMO is stated to be refit on training plus validation data, the grids differ (p.6 l.322-323 vs. p.7 l.352), and no same-protocol run with λR = 0 is reported, while Fig. 5 shows λR moving validation nMSE by about 0.01. The 12 p-values come from naive paired t-tests over five overlapping random splits against "the strongest baseline" (p.6 l.307-311). After the resampled-split correction (Nadeau & Bengio, 2003) and Holm's adjustment, at most one (PD H&Y nMSE) stays below 0.05, and the overall gain of 0.018 nMSE is 0.88 between-split SDs. Given weakness 1, this is why I recommend clear rejection: the evidence does not separate DeMMO's gain over cFSGL from ridge shrinkage, refitting and split noise, so the relation term, which is what DeMMO adds to cFSGL, has no demonstrated benefit.

Questions for authors
  1. Validation experiment: What are the test results of the stage-1 model (λR = 0) under the same refit protocol?
  2. Clarification: Which baseline is each row's comparator, and do the conclusions survive a corrected resampled t-test?
  3. Clarification: What does the relation term add beyond AutoTR's, other than the task axis, symmetry and the ℓ2 penalty?
Reviewer_2DUKOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DeMMO addresses concurrent clinical-status estimation from wearable-derived digital mobility outcomes across disease cohorts and repeated visits. It combines visit-specific linear predictors, temporal regularization, sparse feature selection, and a learned signed relation matrix between longitudinal disease–outcome mappings. This enables information sharing across participant-disjoint cohorts without requiring paired observations across diseases.

The evaluation uses 24 mobility outcomes, five clinical targets, and five visits from four Mobilise-D disease cohorts, comprising 2,204 participants and 7,543 distinct participant–visit records. Experiments use five matched participant-level 70%/10%/20% training/validation/test splits. DeMMO achieves the best reported aggregate predictive performance, reducing overall nMSE from the strongest baseline value of 0.787 to 0.769. Stability-selection analyses identify candidate mobility measures for subsequent clinical investigation.

Critical strengths

The problem is meaningful, and the model is appropriately structured for it. Separating temporal regularity within outcomes from relationships between disease–outcome mappings provides an interpretable approach to sharing information across distinct cohorts. The model accommodates incomplete longitudinal observations without requiring complete trajectories or paired participants across diseases. This is a useful contribution for analysing heterogeneous clinical mobility data.

The internal-validation results are consistently favorable. All visits from a participant remain in the same split, preprocessing uses training data, and evaluation is repeated across five matched splits. DeMMO improves both aggregate metrics for all five outcomes and achieves the lowest mean RMSE in 21 of 25 outcome–visit combinations. Although the average improvement is modest, its consistency supports the practical value of the approach.

Critical weaknesses

The central relation-learning contribution is not sufficiently isolated. The comparison with cFSGL changes both the relation component and the regularization search range. Moreover, setting the relation matrix to zero while retaining a positive relation penalty introduces additional ridge shrinkage, providing a relevant control for determining whether learned relationships contribute beyond stronger regularization. A within-disease-only relation model would further distinguish cross-disease sharing from sharing between outcomes of the same disease. Given the approximately 2.3% aggregate nMSE improvement, these controls materially affect the interpretation of the main technical contribution.

Questions for authors

Could you clarify whether the gains remain when the shared regularization parameters are tuned over identical ranges? A targeted comparison with a ridge-regularized model without learned relations and a within-disease-only relation model would help establish whether cross-disease connections provide an additional benefit. If these comparisons are unavailable, please clarify what the current evidence can establish about this contribution.

Could you clarify which aspect of the relation-learning formulation is methodologically new relative to the closest existing parameter-based multitask approaches? In particular, it would help to distinguish novelty in the learning formulation from the contribution of addressing this clinical modelling setting.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.