A Dual-Use Framework for Clinical Gait Analysis: Attention-Based Sensor Optimization and Automated Dataset Auditing

TL;DR

The paper tackles quantitative gait analysis with wearable IMUs by addressing two core challenges: identifying minimal, task-specific sensor configurations and auditing datasets for hidden biases. It introduces an interpretable, attention-based multi-stream CNN that learns sensor-specific features and outputs cohort-level attention weights via $\alpha_i = \frac{\exp(e_i)}{\sum_j \exp(e_j)}$ and a context vector $c = \sum_i \alpha_i v_i$ to drive binary classification. Applied to Voisard et al.'s multi-cohort gait dataset across four tasks (PD screening, OA screening, CVA asymmetry, PD vs CVA differential), the framework reveals a severe right-foot laterality confound in OA and CVA cohorts while providing data-driven sensor synergies for optimized protocols. Beyond sensor optimization, the method functions as an automated data auditor, flagging hidden confounds and guiding future dataset design to improve robustness and clinical applicability of gait analyses.

Abstract

Objective gait analysis using wearable sensors and AI is critical for managing neurological and orthopedic conditions. However, models are vulnerable to hidden dataset biases, and task-specific sensor optimization remains a challenge. We propose a multi-stream attention-based deep learning framework that functions as both a sensor optimizer and an automated data auditor. Applied to the Voisard et al. (2025) multi-cohort gait dataset on four clinical tasks (PD, OA, CVA screening; PD vs CVA differential), the model's attention mechanism quantitatively discovered a severe dataset confound. For OA and CVA screening, tasks where bilateral assessment is clinically essential, the model assigned more than 70 percent attention to the Right Foot while statistically ignoring the Left Foot (less than 0.1 percent attention, 95 percent CI [0.0-0.1]). This was not a clinical finding but a direct reflection of a severe laterality bias (for example, 15 of 15 right-sided OA) in the public dataset. The primary contribution of this work is methodological, demonstrating that an interpretable framework can automatically audit dataset integrity. As a secondary finding, the model proposes novel, data-driven sensor synergies (for example, Head plus Foot for PD screening) as hypotheses for future optimized protocols.

Figures (4)

• Figure 1: Mean Sensor Attention Weights for Parkinson's Disease (PD) Screening. The model learned to prioritize a novel synergy between the Right Foot (RF) (52.5%) and Head (HE) (32.9%) sensors. This data-driven hypothesis aligns with the clinical presentation of PD, which involves both appendicular (gait) and axial (postural) motor deficits. Error bars represent 95% confidence intervals calculated from 1000 bootstrap samples, indicating the statistical uncertainty in the estimate of the mean attention.
• Figure 2: Mean Sensor Attention Weights for Osteoarthritis (OA) Screening. This result demonstrates the model's function as a data auditor. It assigned over 70% of its attention to the Right Foot (RF) sensor while statistically ignoring the Left Foot (LF) (0.1%). This counter-intuitive pattern reflects the model's discovery of a dataset confound (a 15/0 right-sided laterality bias in the HOA cohort), not a generalizable clinical signature of the disease. Error bars represent 95% confidence intervals (1000 bootstrap samples).
• Figure 3: Mean Sensor Attention Weights for Asymmetry Detection (CVA). Similar to the OA task, the model relied almost exclusively on the Right Foot (RF) sensor (77.4%). The lack of attention to the Left Foot (LF) (0.0%) for a task predicated on asymmetry demonstrates the exploitation of a dataset confounder (a 47/2/0 right-dominant laterality bias in the CVA cohort), validating the attention mechanism's ability to flag hidden biases. Error bars represent 95% confidence intervals (1000 bootstrap samples).
• Figure 4: Mean Sensor Attention Weights for Differential Diagnosis (PD vs. CVA). To solve this complex task, the model learned a sophisticated synergy, allocating importance to the Head (HE) sensor (51.5%) and the Left Foot (LF) sensor (46.8%). It learned to ignore the confounded Right Foot sensor (1.5%), suggesting a strategy of comparing central postural control (Head) with a non-confounded limb (Left Foot). Error bars represent 95% confidence intervals (1000 bootstrap samples).