Sit Posture Classification using XGBoost Algorithm Based on Wearable Sensor Data

Prolonged non-ergonomic sitting posture during computer-based learning activities poses a risk of musculoskeletal disorders, yet current posture monitoring systems are largely static and neglect temporal aspects. This study develops a real-time posture classification system integrating accelerometer-based wearables with the XGBoost algorithm, specifically designed for university students. Data were collected from 21 subjects using two MPU6050 sensors placed on the thoracic and lumbar areas, yielding 28,054 data rows. Preprocessing involved mean imputation for missing values, removal of non-feature columns, and exclusion of gyroscope data due to extreme outliers. Feature extraction using 30-sample non-overlapping windows generated 24 statistical features (minimum, maximum, mean, standard deviation) from the accelerometer axes. As the primary benchmark for generalization, the XGBoost model was evaluated using Leave-One-Person-Out Cross-Validation (LOPOCV), achieving an average accuracy of 73.76% with balanced precision and recall scores (0.73–0.75), indicating consistent performance when encountering new subjects. Feature importance analysis identified az2_max (score: 0.1087) and az1_mean (score: 0.1053) as the most dominant features. Meanwhile, as an initial feasibility test under controlled conditions, real-time testing with five subjects (50 instances) reached 100% accuracy, with a web monitoring interface successfully displaying posture status, confidence values (53–81%), and posture-change logs, alongside a buzzer mechanism that activated appropriately upon detecting non-ergonomic postures. This system demonstrates the potential of low-cost accelerometer-based wearables for effective real-time posture monitoring; however, its performance is significantly influenced by individual physical characteristics such as height and weight. Consequently, future research is advised to diversify the dataset and implement personalization techniques to enhance generalization across broader populations.