Observable Performance Does Not Fully Reflect System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
| Authors | Jacques Raynal et al. |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2605.00778 |
| Download | |
| Categories | cs.LG |
Abstract
In biomechanical systems, observable performance is often used as a proxy for underlying system organization. However, this assumption implicitly presumes a correspondence between output metrics and internal system states that may not hold in adaptive systems. In this study, the vertical dimension of occlusion (VDO) is considered as a constraint applied to an adaptive neuromechanical system, enabling the exploration of system-level responses under controlled variations. A single-case design in a patient with Parkinson's disease allows an intra-individual analysis across repeated conditions.The analysis is structured across three complementary levels: (i) aggregated linear metrics describing observable performance, (ii) a dynamical systems framework describing temporal organization in state space, and (iii) a latent space representation obtained through unsupervised embedding. The results show that conditions with comparable observable performance may correspond to different organizations in both state space and latent space representations. This dissociation highlights a limitation of aggregated metrics and suggests that similar outputs may arise from non-equivalent system states. A fourth level is proposed as a purely conceptual extension describing potential relationships between system states. This level is not implemented and is not derived from experimental data. These observations are strictly exploratory and non-causal. The proposed framework does not establish mechanistic, predictive, or directional relationships, but provides a structured approach for analyzing constraint-driven systems across multiple levels of representation.
Engineering Breakdown
Plain English
This paper investigates how observable performance metrics in adaptive neuromechanical systems (like patients with Parkinson's disease) may not accurately reflect the underlying internal organization of those systems. The researchers used vertical dimension of occlusion (VDO) as a controlled constraint and analyzed a single patient across repeated conditions using three complementary analytical levels: linear aggregate metrics, dynamical systems analysis, and (implied from the abstract structure) deeper state-space characterization. The key finding is that output performance alone can be misleading—the same external metrics can emerge from fundamentally different internal system organizations, especially in adaptive systems that can compensate and reorganize. This challenges a widespread assumption in biomechanics that performance metrics directly map to system state.
Core Technical Contribution
The paper's core novelty is demonstrating that in adaptive biological systems, the input-output relationship is not bijective: multiple distinct internal system states can produce identical or nearly identical observable performance. Rather than treating performance metrics as direct windows into system organization, the authors propose a three-level hierarchical analysis framework that combines traditional linear metrics with nonlinear dynamical systems analysis to disambiguate different mechanisms that produce the same output. This is methodologically significant because most biomechanical and neurological research assumes performance metrics directly reflect system health or organization; this work shows that assumption breaks down under adaptation. The contribution is both conceptual (challenging a false equivalence) and methodological (providing a framework to detect it).
How It Works
The analysis operates at three complementary levels of abstraction. First, aggregated linear metrics (classical kinematic/kinetic measures like range of motion, force profiles, timing parameters) describe what an observer can measure directly—the 'output' of the system. Second, a dynamical systems framework reconstructs the temporal organization and state-space structure from time-series data, revealing whether the system exhibits stable attractors, chaos, bifurcations, or compensatory reorganization over the repeated conditions. Third (implied from the abstract's structure), deeper state-space analysis likely examines the dimensionality and structure of the neural-mechanical coupling, possibly using techniques like embedding dimension estimation or Lyapunov exponents. The input is the VDO constraint (a biomechanical parameter that changes across conditions), and the output is characterization of whether the system reorganizes internally while maintaining similar external performance, revealing the distinction between surface-level similarity and mechanistic equivalence.
Production Impact
For engineers building diagnostic or monitoring systems in neuromechanics, rehabilitation, or adaptive prosthetics, this work has critical implications: relying solely on performance metrics (e.g., 'patient achieved 95% range of motion') can mask deterioration or functional reorganization of the underlying system. In production systems, you would need to integrate multi-level analysis pipelines: first compute traditional metrics for real-time feedback, then periodically run dynamical systems analysis (embedding, attractor reconstruction, entropy measures) on longer time windows to detect whether the system is compensating or degrading. This increases computational cost modestly (embedding reconstruction is O(n log n) for n time steps) and requires longer observation periods (multiple cycles to build sufficient statistics), but prevents false reassurance from stable metrics masking dangerous internal reorganization. For clinical decision support, this means augmenting performance dashboards with state-space visualizations and stability indices so clinicians see both what the patient can do and how the patient's nervous system is organizing itself.
Limitations and When Not to Use This
The paper is constrained to a single-case design (one Parkinson's patient), which severely limits generalizability—it proves the concept that performance-state decoupling can occur, but does not establish how often it occurs, in what conditions, or across what populations. The VDO constraint is biomechanically specific; it's unclear whether the three-level framework generalizes to other adaptive systems (e.g., cognitive tasks, motor learning in younger populations, or systems without neurodegenerative disease). The paper does not appear to propose automated detection rules—a clinician or engineer still needs to visually interpret state-space plots and dynamical metrics, which requires expertise and is not scalable to large patient cohorts. Finally, the abstract truncates before describing the state-space analysis level in detail, leaving ambiguity about which specific dynamical systems techniques were used and their robustness to noise, non-stationarity, or parameter sensitivity.
Research Context
This work builds on decades of research in motor control that distinguishes task-level goals from internal neuromuscular organization (Bernstein's degrees-of-freedom problem, motor redundancy literature). It extends dynamical systems approaches in neuroscience (e.g., work on neural manifolds and latent dynamics) into the biomechanical domain and challenges the assumption underlying many biomarker-development studies that observable metrics like gait speed or force production directly index system health. The paper aligns with recent emphasis on temporal dynamics and state-space analysis in movement neuroscience, moving away from snapshot metrics toward time-series characterization. It opens research directions in multi-level system identification for neuromechanics, differential diagnosis (can we detect disease type from state-space topology rather than metric values?), and adaptive system theory applied to rehabilitation.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::