Prespecified held-out in silico robustness validation of a plantar-pressure gait-event classifier under sensor and transport perturbations

Abstract

Wearable plantar-pressure signal processing for gait-event classification must remain robust when signals are affected by noise, drift, channel mismatch, data loss, or non-walking load cycles. We compared the NUWAC gait-event and plantar-contact classifier R4 with frozen R3 in a prespecified held-out in silico validation against synthetic latent truth. Both algorithms were frozen before unseen fixtures were generated once, locked, and processed through the exact RAW20 path. We evaluated 31 positive conditions (310,000 contacts per algorithm), nine negative controls (54,000 synthetic seconds per algorithm), and temporal resolution. R4 met all six superiority criteria and all prespecified engineering guardrails. R4-minus-R3 improvements were 23.47 percentage points for the prespecified classification score at 5% noise, 14.00 points for detection F1 at 10% noise, 42.05 points for the prespecified classification score at 10% noise, 52.00 points for recall at 0.3%/s drift, 8.21 points for recall at offset SD 25 counts, and 24.01 points for the prespecified classification score under combined-moderate stress. R4 produced no false steps across the negative controls but made later decisions under noise and combined stress. Residual boundaries included 20% noise, large offset mismatch, noHeel contacts under severe perturbation, and a small recall loss under 20% gain variation. These findings establish software-level behavior under specified computational conditions; they do not constitute physical-sensor, human reference-system, device, or clinical validation. Version 2 note (29 August 2026): This version aligns the preprint with the manuscript submitted to Medical Engineering & Physics. It updates the title and scientific terminology; clarifies the preserved replicate-mean classification score versus strict pooled matched-event classification accuracy; incorporates paired Monte Carlo uncertainty and the post hoc ±100-ms matching-window sensitivity analysis; and adds author ORCID iDs. The frozen R3 and R4 algorithms, held-out fixtures, prespecified primary endpoints, and held-out validation results were not regenerated or altered for this revision. Preprint status: This manuscript was submitted to Medical Engineering & Physics on 29 August 2026 and has not yet undergone peer review. Authorship note: Omid Aslani Damirchi and Hamidreza Asgari contributed equally to this work.