Bilateral Ground Reaction Force Prediction Using Deep Learning Models and Custom Force Plate

Several low-cost force plates have been proposed as alternatives for laboratory-grade force plates. Nevertheless, the inability to quantify bilateral ground reaction force (GRF) prevents these inexpensive force plates from being used for biomechanical analysis and certain clinical metric acquisition. This study developed deep-learning models, such as autoencoder and U-net, to predict bilateral GRF from vertical GRF measured using a low-cost custom force plate during sit-to-stand, gait initialization, and gait. Results indicated that the U-net model, which utilized STFT vertical GRF as input, performed the best. In addition to predicting the mediolateral GRF measured during sit-to-stand, the model accurately predicted the anterior-posterior and mediolateral GRF for sit-to-stand, gait initialization, and gait in the test dataset, achieving high Pearson's correlation coefficient, coefficient of determination, and intraclass correlation coefficient values of over 0.90, 0.79, and 0.89, respectively. The model demonstrated a higher Pearson's correlation coefficient compared to three related previous studies that utilized different methods to predict anterior-posterior GRF and six studies in inferring mediolateral GRF. The results demonstrated the potential of TFU and custom force plate as a GRF measurement tool to perform bio-mechanical analysis. ABSTRAK: Beberapa plat daya kos rendah telah dicadangkan sebagai alternatif kepada plat daya berkualiti makmal. Walau bagaimanapun, ketidakmampuan untuk mengukur daya reaksi tanah (GRF) secara bilateral menghalang plat daya yang murah ini daripada digunakan untuk analisis biomekanik dan pengambilan metrik klinikal tertentu. Kajian ini membangunkan model pembelajaran mendalam, seperti autoencoder dan U-net, untuk meramalkan GRF bilateral daripada GRF menegak yang diukur menggunakan plat daya khas kos rendah semasa pergerakan duduk-ke-berdiri, permulaan berjalan, dan berjalan. Hasil menunjukkan bahawa model U-net, yang menggunakan GRF menegak STFT sebagai input, memberikan prestasi terbaik. Selain meramalkan GRF mediolateral yang diukur semasa duduk-ke-berdiri, model ini juga meramalkan dengan tepat GRF anterior-posterior dan mediolateral untuk duduk-ke-berdiri, permulaan berjalan, dan berjalan dalam set data ujian, mencapai nilai koefisien korelasi Pearson, koefisien penentuan, dan koefisien korelasi intrakelas yang tinggi melebihi 0.90, 0.79, dan 0.89, masing-masing. Model ini menunjukkan koefisien korelasi Pearson yang