AI predicts gait changes from jaw position in Parkinson's study
Neural network models how your bite affects walking patterns.
A new arXiv preprint (2605.15862) from Jacques Raynal and colleagues explores whether machine learning can approximate the internal dynamics of gait under different jaw positions—a concept known as occlusal constraint. Using a single-subject design with a Parkinsonian participant, the team recorded walking patterns with instrumented insoles across two sessions 11 weeks apart. They tested six observational probes: natural occlusion, open-mouth disengagement, strong clenching, and three vertical-dimension changes in centric relation.
The researchers applied Principal Component Analysis (PCA) to build a latent-space representation (PC1–PC2) of gait, then trained a feed-forward neural network to approximate the observed longitudinal transformation (M1→M2). The model successfully preserved the displacement hierarchy among the three centric-relation conditions (OC3 < ONL < OC2.5) and maintained the global structure in a six-probe analysis. However, the authors emphasize this is not a predictive clinical tool—it's a methodological bridge showing that observed latent transformations can be internally approximated within a single dataset.
- Single-subject study on a Parkinson's patient using instrumented insoles and 6 jaw occlusion conditions
- PCA latent-space model + feed-forward neural network preserved gait transformation ordering across 11 weeks
- Authors explicitly state no clinical prediction, causal effects, or generalizability—purely methodological
Why It Matters
Bite and gait connection could unlock new non-invasive diagnostics via latent-space ML models.