Complexity and Stability of Neural Activity Across Aging and Neurodegenerative Disease

The paper Complexity and Stability of Neural Activity Across Aging and Neurodegenerative Disease proposes a distributional and geometric framework to characterize the temporal stability of neural population activity using EEG. The authors model EEG signals as empirical distributions of windowed activity patterns and quantify temporal stability using Wasserstein distance, while intrinsic dimensionality captures representational complexity.

Objective and Methods:

The study addresses whether the brain reuses similar activity patterns over time within a fixed cognitive state. The authors model EEG as distributions of windowed activity patterns and quantify temporal stability using Wasserstein distance, while intrinsic dimensionality captures representational complexity. They analyzed three complementary EEG datasets: a multi-task dataset (60 healthy university students, ages 18–28), a lifespan dataset (365 healthy participants across a wide age range), and a clinical dataset (374 participants: 136 healthy controls, 129 with mild cognitive impairment, and 109 with Alzheimer’s disease). EEG signals were divided into non-overlapping 1-second windows, each represented as a 100-dimensional vector. Intrinsic dimensionality was estimated using the maximum likelihood estimator of Levina and Bickel (2004) with neighborhood size K = 50, and Wasserstein distances were computed using the POT package.

Key Findings:

1. Constrained distributional evolution: Neural representations under fixed cognitive conditions show bounded, non-monotonic fluctuations over time rather than progressive drift. Within-condition Wasserstein distances were consistently smaller than cross-condition distances, indicating condition-specific stability. The paper states: Neural representations exhibit constrained distributional evolution within fixed cognitive conditions and neural activity remains within condition-specific regions of representational space without evidence of progressive drift.

2. Spatial organization: Both intrinsic dimensionality and within-condition Wasserstein distance exhibit reproducible posterior–anterior gradients across cognitive conditions. The paper notes: Across conditions, intrinsic dimensionality showed a robust posterior–anterior gradient, with occipital and temporal regions showing higher values and frontal regions showing markedly lower values. A similar pattern was observed for Wasserstein distance, with posterior channels showing larger representational displacement.

3. Healthy aging: Healthy aging is characterized by representational expansion. The paper states: Healthy aging is associated with representational expansion and reduced distributional stability. Both intrinsic dimensionality and within-condition Wasserstein distance increased significantly with age, and these measures were significantly positively correlated across channels after FDR correction.

4. Neurodegenerative disease: Mild cognitive impairment and Alzheimer’s disease show a joint collapse of both measures. The paper states: Neurodegenerative disease is associated with collapse and rigidification of representational geometry. Both clinical groups exhibited lower intrinsic dimensionality and reduced within-condition Wasserstein distance relative to healthy controls. The paper emphasizes: "Reduced within-condition Wasserstein distance in neurodegenerative disease should not be interpreted as preserved functional stability. Rather, in the context of collapsed intrinsic dimensionality, it reflects pathological over-stability."

5. Unified framework: The paper proposes that healthy aging and neurodegeneration correspond to opposite shifts along a shared geometric axis. In healthy aging, neural representations expand into a higher-dimensional and less compact regime. In contrast, neurodegenerative disease is characterized by collapse and rigidification of representational geometry.

Theoretical foundation:

The paper provides mathematical foundations in Appendix A, showing that the expected Wasserstein distance between empirical distributions is controlled by intrinsic dimensionality (Theorem Appendix A.1, based on Weed and Bach, 2019). This provides formal justification for interpreting distributional stability as a geometric property.

Conclusions:

The paper concludes: "Our findings identify intrinsic dimensionality as a unifying geometric principle for understanding the temporal stability of neural representations. By shifting the focus from individual signals to evolving empirical distributions, this work provides a compact framework for understanding how neural population activity remains stable without becoming fixed, and how this balance is systematically altered across cognition, healthy aging, and neurodegenerative disease."

Limitations acknowledged:

The authors note limitations including EEG's limited spatial resolution, the use of statistical proxies for intrinsic dimensionality, and the correlational nature of the study. They suggest future work using source-resolved imaging, invasive electrophysiology, or multimodal recordings, and causal perturbation studies.

Significance:

The paper states: This framework offers a principled approach to quantifying neural representational stability, with potential utility as a sensitive biomarker for tracking cognitive aging and neurodegeneration in clinical settings.

Improvements to AI Systems:

1. Dynamic Stability-Aware Representation Learning: Incorporate Wasserstein-distance-based temporal stability constraints into contrastive or self-supervised learning objectives. This would enable AI models to learn representations that maintain bounded, condition-specific distributions over time, preventing catastrophic drift in continual learning or online adaptation scenarios.

2. Intrinsic Dimensionality as a Regularization Signal: Use the maximum likelihood estimator of intrinsic dimensionality (Levina & Bickel, 2004) as a dynamic regularizer during training. The AI system could automatically adjust its representational complexity—expanding dimensionality for novel tasks and collapsing it for overfitting—mimicking the healthy aging vs. neurodegeneration axis.

3. Geometric Biomarker for Anomaly Detection: Build an AI diagnostic tool that computes per-channel intrinsic dimensionality and within-condition Wasserstein distance from EEG or other neural signals. The system would classify healthy aging (expansion) vs. neurodegenerative disease (collapse) by detecting the joint shift along the shared geometric axis, with FDR-corrected correlations across channels.

4. Posterior–Anterior Gradient Modeling for Spatial AI: Implement a spatial prior in neural network architectures (e.g., for EEG or fMRI analysis) that enforces a reproducible posterior–anterior gradient in representational complexity. This would improve transfer learning across subjects and conditions by aligning learned features to known neuroanatomical gradients.

5. Empirical Distribution-Based Time-Series Encoders: Replace point-wise time-series encoders with distributional encoders that model each window as an empirical distribution and compute Wasserstein distances between windows. This would make AI systems robust to non-stationarity and noise in physiological or financial time-series data, as the framework explicitly handles bounded, non-monotonic fluctuations.

6. Pathological Over-Stability Detection: Develop an AI alert system for clinical monitoring that flags reduced within-condition Wasserstein distance in conjunction with collapsed intrinsic dimensionality—not as a sign of stability but as a marker of pathological rigidification. This prevents false positives in adaptive AI systems deployed for patient monitoring.

7. Unified Cognitive-Aging Axis for Transfer Learning: Train a meta-learning algorithm on the proposed healthy-aging vs. neurodegeneration axis to modulate model capacity. The AI could automatically expand its representational space for underfitting regimes (analogous to healthy aging) and collapse it for overfitting or rigid regimes (analogous to disease), improving generalization across diverse data distributions.

8. Condition-Specific Region Mapping for Interpretable AI: Use the posterior–anterior gradient findings to build explainable AI that attributes model decisions to specific neural channels based on their intrinsic dimensionality and stability. This would enhance interpretability in brain-computer interfaces and neurofeedback systems.

What the Improved AI System Can Do:

• Adaptively balance representational stability and flexibility in real-time learning, avoiding both catastrophic forgetting and pathological over-stability.

• Diagnose and track cognitive decline from EEG with a single geometric framework, distinguishing healthy aging from MCI/Alzheimer’s with higher sensitivity than raw signal features.

• Transfer learned representations across subjects and tasks by aligning to universal posterior–anterior complexity gradients, reducing calibration time in BCI applications.

• Provide interpretable, channel-level explanations of neural state changes, linking AI decisions to biologically meaningful geometric shifts.

• Monitor patients for pathological rigidification without false alarms, using the joint collapse signature rather than naive stability metrics.

• Optimize model capacity dynamically based on the intrinsic dimensionality of incoming data, improving robustness in non-stationary environments (e.g., autonomous systems, financial forecasting).