Gait parameters and plantar pressure estimation from the prosthesis evaluation questionnaire using a synthetic-data-enhanced supervised learning model: A proof-of-concept study

Abstract

Understanding the impact of lower limb prostheses on gait is crucial for improving users’ quality of life. To this end, quantitative gait analysis provides objective and detailed information but is time-consuming and resource-intensive, whereas qualitative usability and ergonomics questionnaires are easier to collect but inherently subjective. To bridge the gap between these modalities, this study proposes a proof-of-concept two-stage machine learning framework that enables the prediction of quantitative gait-related signals from qualitative clinical assessments, particularly in settings with limited data. The proposed approach combines Gaussian Mixture Model (GMM)-based synthetic data generation with Extreme Gradient Boosting (XGB) regression to enhance predictive performance. The dataset consists of motion analysis and plantar pressure measurement experiments collected from 10 transfemoral amputees using six different prosthetic devices, along with corresponding Prosthesis Evaluation Questionnaire (PEQ) responses. The GMM is first employed to increase the effective signal density through controlled synthetic data generation, after which the XGB regressor is trained to estimate biomechanical gait parameters and plantar pressure features directly from questionnaire inputs. The results demonstrate that meaningful quantitative predictions can be achieved from qualitative survey data, highlighting the potential of the proposed framework as a clinically relevant and data-efficient approach for gait assessment. This work serves as an initial proof-of-concept for integrating subjective and objective data modalities within a unified biomedical signal analysis under limited data availability.