Neural network-based prediction of missing key features in vertical GRF-time recordings.

Area of Science:

• Biomechanics
• Biomedical Engineering
• Artificial Intelligence

Background:

• Ground reaction force (GRF) data loss due to faulty force platform stepping is common.
• This data loss necessitates repeat data collection, impacting research efficiency.
• Accurate gait parameter estimation is crucial for clinical and research applications.

Purpose of the Study:

• To propose and validate an artificial neural network (ANN) approach for estimating lost gait parameters from corrupted GRF data.
• To assess the accuracy of ANN models in predicting stance time (ST), push-off force (Fmax2), and push-off time (Tmax2).
• To evaluate the feasibility of reconstructing gait forces from incomplete GRF-time traces.

Main Methods:

• Collected GRF-time data during normal walking from young and elderly individuals.
• Developed back-propagation neural network models using unaffected vertical GRF features as inputs.
• Trained models to predict affected gait features: stance time (ST), push-off force (Fmax2), and push-off time (Tmax2).
• Validated model performance on new gait trial data.

Main Results:

• ANN models accurately predicted missing stance time (ST) with 96.5% accuracy (r>0.9).
• Push-off force (Fmax2) and push-off time (Tmax2) were reconstructed with high accuracy (95.7% and 97.6%, respectively).
• ST prediction accuracy decreased when push-off force/time data were unavailable for the model.

Conclusions:

• ANNs offer a promising method for estimating missing gait parameters (ST, Fmax2, Tmax2) from affected GRF-time traces.
• This approach can effectively reconstruct gait forces from corrupted data, reducing the need for re-acquisition.
• The study highlights the potential of AI in overcoming data integrity issues in biomechanical analysis.