Automatic Sitting Posture Recognition in Wheelchair Users: An AI-driven Pilot Study

Automated recognition of sitting postures in wheelchair users is crucial for preventing pressure ulcers and improving overall comfort and health. This pilot study explores using XGBoost for automatic sitting posture recognition in wheelchair users. XGBoost was chosen for its robustness, ability to handle imbalanced datasets, and its strategy of combining multiple weak learners to create a strong learner, thereby enhancing model accuracy and robustness. The proposed approach is evaluated using a publicly available dataset. Pressure sensors on the seat and an ultrasonic sensor on the backrest were utilized to classify four sitting postures: normal sitting, higher pressure on the right side, higher pressure on the left side, and higher forward pressure. XGBoost demonstrated superior performance with perfect accuracy and F1-score, outperforming Support Vector Machines (SVM) and Naive Bayes classifiers. Additionally, SHAP (SHapley Additive exPlanations) analysis is employed to model explainability and identify the most significant variables for detecting abnormal postures. While results are promising, the limited dataset suggests that further research with more extensive data and additional sensors could enhance accuracy.