If the first post argued that gait can carry early monitoring signal, this one is the workshop floor.
We will walk through how we turn a public foot-force corpus into training examples for a model that flags pattern differences for further review. This is research and engineering practice. It is not a diagnostic product claim, and it is not FDA clearance.
The audience split is intentional:
- Medical readers: what the data actually is, and what a model output should mean clinically
- Gait people: sensors, sampling, dual-task conditions, and why windows beat a single idealized cycle
- AI researchers: leakage, scaling, domain shift, and evaluation that does not flatter itself
The public corpus: PhysioNet Gait in Parkinson’s Disease
We train on Gait in Parkinson’s Disease on PhysioNet (Hausdorff and collaborators; v1.0.0). It aggregates multichannel underfoot force recordings from three published studies:
- Yogev et al. (dual tasking in PD)
- Hausdorff et al. (rhythmic auditory stimulation in PD)
- Frenkel-Toledo et al. (treadmill walking in PD)
The release includes records from 93 people with idiopathic Parkinson’s disease and 73 healthy controls. Subjects walked at a self-selected pace on level ground for about two minutes. Under each foot, eight sensors measured vertical ground reaction force (VGRF) in newtons. Sixteen sensor channels plus left and right totals were sampled at 100 Hz.
File names encode study (Ga / Ju / Si), group (Co control vs Pt patient), subject id, and walk condition. In the Ga study, for example, usual walk is often 01 and dual-task walk (serial-7 subtraction) is 10. Demographics and clinical severity measures (Hoehn and Yahr, UPDRS where available) ship with the release.
Researchers should follow PhysioNet citation and license terms on the project page. We treat the corpus as a research training and benchmarking resource, not as proof that any commercial product is cleared on this data.
Why this corpus is still worth learning from
From a gait science view, the dataset is valuable because it preserves time-resolved force channels, not only summary features. You can rebuild stance and swing structure, approximate center-of-pressure under each foot, and study stride-to-stride dynamics.
From an ML view, it is valuable because it is hard in the right ways:
- Strong person identity in the signal
- Session and condition effects (usual vs dual-task)
- Class imbalance and severity heterogeneity inside PD
- Hardware geometry that is related to, but not identical with, dense modern FSR insoles
That last point is the domain-shift elephant. Public PD force data is not a drop-in match for every commercial insole layout. It is still one of the best openly documented longitudinal foot-force resources for learning gait dynamics under disciplined validation.
What each row looks like
Each line in a walk file has 19 columns:
- time in seconds
- eight left-foot VGRF channels
- eight right-foot VGRF channels
- total left force
- total right force
The documentation’s sensor (X, Y) layout supports center-of-pressure proxies under each foot. If you have spent time staring at plantar heatmaps, the spirit is familiar even when the sensor count differs from a 40-channel adult insole map.
Design choice 1: multi-step windows, not only one gait cycle
Many classical pipelines reduce gait to a single normalized cycle and a handful of scalars. That is clean. It also throws away how cycles chain.
We pass short windows of consecutive steps into the model so it can learn sequence structure: stride-to-stride variability, dual-task disruption, and loading transitions that do not live inside one idealized cycle.
Intuition for clinicians: impairment often shows up in the unfolding of gait, not only in the average of isolated strides. Intuition for ML: windows trade some label purity for temporal context. That trade should be reported whenever results are discussed, because it changes what “accuracy” even means.
Design choice 2: per-session scaling
Absolute force amplitude is entangled with body mass, footwear, sensor calibration, and walking speed. If you feed raw newtons naively, models can cheat on identity or hardware quirks.
We scale features per session so learning focuses more on pattern shape and relative loading. This is also a domain-shift hedge when moving between public force sensors and other insole geometries later.
For gait enthusiasts: think of it as protecting the shape of the loading waveform rather than worshipping absolute magnitude. For medical readers: this is one reason a “score” from force ML should never be read like a lab value with a universal clinical cutpoint unless that cutpoint is separately validated.
Design choice 3: subject-wise splits and strict cross-validation
Gait is highly individual. If the same person appears in both train and test, published numbers look better than real-world generalization. That mistake is still common in wearable ML papers.
We enforce subject-wise partitions: no subject in both training and evaluation folds. Cross-validation is aimed at estimating generalization across people, not across random windows from the same walker.
When dual-task walks exist for a subject, leakage rules stay consistent at the subject level. Condition labels can be analysis slices. They are not a backdoor for identity leakage.
If you work in AI research, this is the non-negotiable section. Everything else is modeling taste. This is validity.
What the model is for (and what it is not)
Trained this way, the model maps windowed force patterns toward research labels and monitoring-oriented pattern flags derived from the corpus (for example, differences associated with PD vs control cohorts in this dataset). Downstream review tools may surface anomalies or longitudinal change for a clinician to inspect.
It is not:
- A diagnosis of Parkinson’s disease or any neurodegenerative disease
- A treatment recommendation
- An FDA-cleared screening device
- A promise that a home walk replaces neurological assessment
Prefer phrasing like: may help surface or flag gait patterns for further clinical review.
We are not publishing specific discrimination metrics in this post. If numbers are approved later, they belong in a separate results note with methodology, confidence intervals, failure modes, and limits.
From public corpus to an analysis path
In our stack, trained models sit behind an internal analysis path that accepts gait sessions and returns structured insights for review tools. Implementation details and commercial packaging are out of scope here.
The research claim is narrower and, we think, more useful: public PD force data, windowed learning, per-session scaling, and subject-wise validation are a defensible way to build gait pattern models without pretending the clinic hallway is the whole world.
Closing
If you work on gait ML, tell us how you handle subject leakage and the cycle-versus-window tradeoff. If you work in clinic, tell us which pattern flags would actually change what you do next visit.
Part A covers why we care about gait as an early monitoring signal: Why gait is an early signal in neurodegenerative disease. Follow VRsteps on LinkedIn for related research notes.
