This pilot study aimed to explore using pressure mat sensors and machine learning models—Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU)—to objectively classify wheelchair fencers into categories.
Eight members of the Hong Kong wheelchair fencing team (four male, four female) were recruited from the Sports Association for the Physically Disabled of Hong Kong (HKSAPD) to train machine learning models. Participants, with an average of 5.2 years of competitive experience, were classified into WFC categories A (four athletes) and B (four athletes). Four additional participants, without prior fencing experience but meeting the minimum impairment criteria, were recruited for Part B (validation study). All were classified by a qualified classifier and provided informed consent.
In Part A, participants performed eight functional tests from the WFC classification protocol. Pressure data from mats (Tekscan CONFORMat) were used to assess Centre of Force displacement. Machine learning models (RNN, LSTM, GRU) were trained using 1,067 data files from all participants, and their performance was validated using leave-one-motion-out cross-validation. The most accurate model was selected based on feature importance analysis. The model was then validated with new data in Part B.
The most accurate machine learning model was identified first. RNN, LSTM, and GRU models were trained using all six movement features, and their prediction accuracy for fencer classification was evaluated. The GRU model outperformed the others and was selected for further validation. The average prediction accuracy across all motions was > 0.7, and for individual participants, it exceeded 0.8. The GRU model demonstrated high accuracy in predicting participants' categories.
Using the prediction model from Part A, the average accuracy for the validation participants was 0.90. Notably, bilateral side flexion achieved a perfect prediction accuracy of 1.00 for all participants. Upper back extension and side balance with a weapon also showed strong accuracies of 0.90 and 0.81, respectively.
This proof-of-concept study demonstrated that the GRU model achieved the highest accuracy for wheelchair fencer classification. The prediction model showed moderate to high accuracy, but validation results were limited by variations in training background, skill levels, and a small sample size.
This study highlights the potential of artificial intelligence in enhancing the accuracy of wheelchair fencing classification. Future studies with larger samples and the inclusion of video data may improve results. The same principle can be applied in other disabled sports.
artificial intelligence
machine learning
