Observable Performance Does Not Fully Reflect System Organization: A Multi-Level Analysis of Gait Dynamics Under Occlusal Constraint
Unsupervised embedding reveals distinct internal organizations producing identical walking metrics in Parkinson's patient.
A new preprint (arXiv:2605.00778) by Jacques Raynal, Pierre Slangen, and Jacques Margerit challenges a core assumption in biomechanics and adaptive systems: that observable performance reliably reflects underlying system organization. Using a single-case design with a Parkinson's disease patient, the team manipulated the vertical dimension of occlusion (bite height) as a constraint and analyzed gait across three levels: traditional linear metrics, a dynamical systems framework (state-space topology), and an unsupervised latent space embedding.
Their key finding: different bite heights sometimes produced nearly identical aggregated metrics (e.g., stride length, speed), yet the state-space and latent representations revealed fundamentally different internal organizations. In one example, two conditions with almost identical walking speeds mapped to distinct clusters in the embedding space. This suggests that similar outputs can arise from non-equivalent system states—undermining the common practice of using performance as a proxy for internal robustness. The authors propose a fourth conceptual level for future work but emphasize the results are exploratory and non-causal.
- Conditions with nearly identical walking speed and stride length showed different latent space clusters via unsupervised embedding.
- Three-level analysis: aggregated metrics, dynamical systems state space, and latent embeddings—each revealing different insights.
- Single-case design on a Parkinson's patient allows controlled intra-individual comparison but limits generalization.
Why It Matters
For ML and systems engineering, performance metrics alone can mask critical internal reorganization—a caution for model evaluation and health monitoring.