Improved classification of Alzheimer's disease and mild cognitive impairment through dynamic functional network analysis
Pith reviewed 2026-05-22 17:12 UTC · model grok-4.3
The pith
Dynamic functional connectivity from fMRI time series enables robust classification of Alzheimer's, mild cognitive impairment, and healthy controls.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
While healthy controls and mild cognitive impairment groups show similar node-level static and dynamic patterns, Alzheimer's disease cases exhibit clearer differences: stable connectivity reductions between white-matter regions and parietal/somatosensory cortices, plus temporally varying changes in amygdala and hippocampal connections. Node centrality measures indicate these white-matter effects are mostly local. Training a Random Forest model on regional BOLD time series informed by both static and dynamic metrics produces robust classification of the MCI, AD, and HC groups, demonstrating added value from the temporal dimension.
What carries the argument
Sliding-window dynamic connectivity derived from resting-state fMRI, combined with a Random Forest classifier trained on regional BOLD time series.
If this is right
- Stable static differences appear between white-matter regions and parietal/somatosensory cortices.
- Temporally varying differences concentrate in connections involving the amygdala and hippocampal formation.
- White-matter connectivity differences register as predominantly local when examined through node centrality.
- Combining static and dynamic metrics in the classifier improves separation of the three clinical groups over static information alone.
Where Pith is reading between the lines
- The same dynamic features might be tested as early predictors of conversion from mild cognitive impairment to Alzheimer's in follow-up scans.
- Applying the sliding-window plus Random Forest pipeline to other neurodegenerative cohorts could identify shared or disease-specific temporal signatures.
- If dynamic measures prove stable across scanners, they could be added to existing diagnostic pipelines to reduce reliance on cross-sectional snapshots.
Load-bearing premise
The sliding-window parameters and Juelich atlas partition actually isolate genuine time-varying brain dynamics rather than analysis artifacts, and the matched ADNI-3 groups are representative enough for the classification results to generalize.
What would settle it
Re-running the Random Forest classifier on the same BOLD data using only static connectivity features yields classification accuracy no higher than chance or baseline static-only performance.
read the original abstract
Brain networks from functional MRI have advanced our understanding of cortical activity and its disruption in neurodegenerative disorders. Recent work has increasingly focused on dynamic (time-varying) brain networks that capture both spatial and temporal patterns of regional co-activity, yet this approach remains underexplored across the Alzheimer's disease (AD). We analysed age- and sex-matched static and dynamic functional brain networks derived from resting-state fMRI data in 315 individuals with AD, mild cognitive impairment (MCI), and cognitively normal healthy controls (HC) from the ADNI-3 cohort. Functional networks were constructed using the Juelich brain atlas, with static connectivity estimated from full time series and dynamic connectivity derived using a sliding-window approach. Group differences were assessed at both link and node levels using non-parametric statistics and bootstrap resampling. While HC and MCI exhibited similar static and dynamic patterns at the node level, clearer differences emerged in AD. Stable (stationary) differences in functional connectivity were identified between white matter regions and parietal and somatosensory cortices, whereas temporally varying differences were consistently observed in connections involving the amygdala and hippocampal formation. Node centrality analysis further suggested that white matter connectivity differences are predominantly local in nature. These findings highlight both shared and distinct functional connectivity patterns across static and dynamic networks, underscoring the importance of incorporating temporal dynamics into brain network analyses of the Alzheimer's spectrum. Additionally, a Random Forest model trained on regional BOLD time series informed by static and dynamic metrics achieved robust classification of MCI, AD, and HC groups, demonstrating the diagnostic potential of time-varying connectivity.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript analyzes static and dynamic functional connectivity networks from resting-state fMRI in 315 age- and sex-matched ADNI-3 subjects (AD, MCI, HC) using the Juelich atlas. Static networks are derived from full time series; dynamic networks use a sliding-window approach. Non-parametric statistics and bootstrap resampling identify group differences, with stable white-matter to parietal/somatosensory differences and temporally varying amygdala/hippocampal differences; node centrality suggests predominantly local white-matter effects. A Random Forest classifier trained on regional BOLD time series informed by these static and dynamic metrics is reported to achieve robust classification of the three groups.
Significance. If the classification performance and group-difference findings hold under proper validation, the work would support incorporating dynamic network features to improve differentiation along the Alzheimer's spectrum beyond static connectivity alone. Positive aspects include the matched public ADNI-3 cohort, non-parametric statistics with bootstrap resampling, and examination of both link- and node-level effects. The absence of quantitative performance metrics, however, limits assessment of practical or clinical significance.
major comments (2)
- [Methods] Methods (dynamic network construction): The sliding-window parameters (length, step size, overlap) are unspecified. These choices directly control the temporal scale at which varying connections (e.g., amygdala/hippocampus) are detected and therefore determine the node/link features supplied to the Random Forest; without them the reported group differences and classifier performance cannot be distinguished from window-induced autocorrelation or atlas boundary effects.
- [Abstract] Abstract and Results (classification): The central claim that the Random Forest model 'achieved robust classification' of MCI, AD, and HC is unsupported by any reported accuracy, AUC, sensitivity/specificity, cross-validation procedure, or statistical thresholds. This quantitative gap is load-bearing for the diagnostic-potential conclusion.
minor comments (2)
- [Methods] The exact feature vector supplied to the Random Forest (how regional BOLD time series are combined with static and dynamic metrics) should be stated explicitly, including any dimensionality reduction or selection steps.
- [Methods] The rationale for selecting the Juelich atlas and any sensitivity checks against alternative parcellations should be added to strengthen the node-centrality claims about white-matter locality.
Simulated Author's Rebuttal
We thank the referee for their constructive and detailed review of our manuscript. We have revised the manuscript to address the identified gaps in methodological detail and quantitative reporting. Our point-by-point responses to the major comments follow.
read point-by-point responses
-
Referee: [Methods] Methods (dynamic network construction): The sliding-window parameters (length, step size, overlap) are unspecified. These choices directly control the temporal scale at which varying connections (e.g., amygdala/hippocampus) are detected and therefore determine the node/link features supplied to the Random Forest; without them the reported group differences and classifier performance cannot be distinguished from window-induced autocorrelation or atlas boundary effects.
Authors: We agree that explicit specification of the sliding-window parameters is essential for reproducibility and to allow readers to assess potential influences of temporal scale on the detected dynamics. These parameters were not stated in the Methods section of the submitted manuscript. We have revised the manuscript to provide a full description of the dynamic network construction, including the window length, step size, and overlap, together with the rationale for their selection drawn from the dynamic functional connectivity literature. We have also added a sensitivity analysis demonstrating that the primary group differences (stable white-matter to parietal/somatosensory effects and temporally varying amygdala/hippocampal effects) remain consistent across a range of plausible window parameters. These changes clarify the temporal scale of the analysis and address concerns about autocorrelation or atlas-related artifacts. revision: yes
-
Referee: [Abstract] Abstract and Results (classification): The central claim that the Random Forest model 'achieved robust classification' of MCI, AD, and HC is unsupported by any reported accuracy, AUC, sensitivity/specificity, cross-validation procedure, or statistical thresholds. This quantitative gap is load-bearing for the diagnostic-potential conclusion.
Authors: We acknowledge that the original abstract and results referred to 'robust classification' without supplying the supporting quantitative metrics, cross-validation details, or statistical thresholds. This omission limits evaluation of the practical significance of the findings. We have revised both the abstract and the Results section to include the classification performance measures (accuracy, AUC, sensitivity, and specificity), the cross-validation procedure, and the results of permutation-based statistical testing against chance. These additions directly support the diagnostic-potential interpretation and allow readers to assess the strength of the reported classification. revision: yes
Circularity Check
No significant circularity in empirical network analysis and classification
full rationale
The manuscript describes an empirical pipeline: construction of static and dynamic functional networks from ADNI-3 rs-fMRI data using the Juelich atlas and a sliding-window procedure, followed by non-parametric group comparisons and Random Forest classification on extracted features. No equations, derivations, or first-principles claims are advanced that reduce to their own inputs by construction. The dynamic-network step employs a standard sliding-window technique without any fitted parameter being relabeled as a prediction, and no self-citation chain or uniqueness theorem is invoked to justify core choices. The reported classification performance is therefore an independent empirical outcome evaluated against an external public cohort rather than a tautological restatement of the analysis inputs.
Axiom & Free-Parameter Ledger
free parameters (1)
- sliding window length and step
axioms (1)
- domain assumption Juelich atlas provides a valid parcellation for constructing functional networks in this cohort
Lean theorems connected to this paper
-
IndisputableMonolith/Cost/FunctionalEquation.leanwashburn_uniqueness_aczel unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
dynamic connectivity derived using a sliding-window approach... Δt=20,40,or 60 data points... half-overlapping windows
-
IndisputableMonolith/Foundation/RealityFromDistinction.leanreality_from_one_distinction unclear?
unclearRelation between the paper passage and the cited Recognition theorem.
Random Forest model trained on regional BOLD time series informed by static and dynamic metrics
What do these tags mean?
- matches
- The paper's claim is directly supported by a theorem in the formal canon.
- supports
- The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
- extends
- The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
- uses
- The paper appears to rely on the theorem as machinery.
- contradicts
- The paper's claim conflicts with a theorem or certificate in the canon.
- unclear
- Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.