Robustness and Interpretability of Deep Learning on Raw Gait Marker Trajectories for Osteoarthritis Classification

Gait analysis is widely used to investigate alterations associated with osteoarthritis (OA), typically relying on biomechanical feature extraction and expert-driven preprocessing. Although machine learning approaches for OA detection have shown promising performance, their robustness and interpretability remain critical concerns for practical deployment. In this work, we investigate whether raw three-dimensional marker trajectories acquired with an optoelectronic system are sufficient to discriminate between healthy subjects and individuals with knee OA. Using a dataset of 313 subjects, a convolutional neural network is trained directly on these raw trajectories, bypassing traditional spatiotemporal or kinematic feature engineering. Beyond classification performance, we conduct a robustness analysis by perturbing the data with increasing noise levels, enabling a systematic evaluation of performance degradation that may arise in acquisition settings with lower precision. Model interpretability is explored using SHAP-based feature importance, identifying the spatial and temporal patterns driving the classification and assessing their stability under noise. Our results suggest that raw gait marker trajectories contain sufficient discriminative information for OA classification (with AUC scores in the order of 0.8, higher than that obtained from kinematics trajectories), and that the proposed model tolerates trajectory perturbations modeled as additive Gaussian white noise with a spatial standard deviation up to 10 cm (5 cm effective after averaging the markers of each segment) and temporally correlated noise, while it is more sensitive to cumulative drift. These findings support the viability of deep learning approaches on minimally preprocessed clinical data and indicate a degree of robustness to trajectory perturbations, suggesting the potential of this approach for osteoarthritis evaluation in low-precision or home-based motion capture scenarios.

Authors

Institutions

Publication Details

Journal
Applied Sciences
Published
2026-11-16
DOI
https://doi.org/10.3390/app162211355
Primary Topic
Gait Recognition and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Robustness and Interpretability of Deep Learning on Raw Gait Marker Trajectories for Osteoarthritis Classification

P. Oliva, Alessandra Retico, Gianmarco Tiddia, Francesca Mainas et al.
4 links
Applied Sciences
Gait Recognition and Analysis
article

Robustness and Interpretability of Deep Learning on Raw Gait Marker Trajectories for Osteoarthritis Classification

P. Oliva, Alessandra Retico, Gianmarco Tiddia, Francesca Mainas, Giuseppe Antonio Motisi
article en

Abstract

Gait analysis is widely used to investigate alterations associated with osteoarthritis (OA), typically relying on biomechanical feature extraction and expert-driven preprocessing. Although machine learning approaches for OA detection have shown promising performance, their robustness and interpretability remain critical concerns for practical deployment. In this work, we investigate whether raw three-dimensional marker trajectories acquired with an optoelectronic system are sufficient to discriminate between healthy subjects and individuals with knee OA. Using a dataset of 313 subjects, a convolutional neural network is trained directly on these raw trajectories, bypassing traditional spatiotemporal or kinematic feature engineering. Beyond classification performance, we conduct a robustness analysis by perturbing the data with increasing noise levels, enabling a systematic evaluation of performance degradation that may arise in acquisition settings with lower precision. Model interpretability is explored using SHAP-based feature importance, identifying the spatial and temporal patterns driving the classification and assessing their stability under noise. Our results suggest that raw gait marker trajectories contain sufficient discriminative information for OA classification (with AUC scores in the order of 0.8, higher than that obtained from kinematics trajectories), and that the proposed model tolerates trajectory perturbations modeled as additive Gaussian white noise with a spatial standard deviation up to 10 cm (5 cm effective after averaging the markers of each segment) and temporally correlated noise, while it is more sensitive to cumulative drift. These findings support the viability of deep learning approaches on minimally preprocessed clinical data and indicate a degree of robustness to trajectory perturbations, suggesting the potential of this approach for osteoarthritis evaluation in low-precision or home-based motion capture scenarios.

Applied SciencesVol. 16(22)
University of Pisa (IT), Istituto Nazionale di Fisica Nucleare (IT), Istituto Nazionale di Fisica Nucleare, Sezione di Cagliari (IT), Istituto Nazionale di Fisica Nucleare, Sezione di Pisa (IT)
European Commission, Ministero della Salute, Instituto Nazionale di Fisica Nucleare
Good health and well-being
Gait Recognition and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon