Towards improved human arm movement analysis: Advanced feature engineering and model optimization

Human arm movement (HAM) analysis plays a crucial role in rehabilitation, sports science, and human-computer interaction (HCI). This has led to extensive research, introducing various systems and algorithms to capture and analyze human movement data. This study presents a novel integrated pipeline for analyzing human arm movement time series data collected from an Inertial Measurement Unit (IMU) embedded in an Apple Watch. The pipeline combines time and frequency domain feature extraction, dimensionality reduction using Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE), and embedded feature selection methods including Least Absolute Shrinkage and Selection Operator (Lasso), Random Forest, and Elastic Net, to optimize classification accuracy while reducing computational complexity. Targeting key movements such as elbow extension/flexion, shoulder extension/flexion, and wrist protonation/supination, data were captured using accelerometers and gyroscopes. The initial dataset, consisting of 720 features from both time-domain and frequency-domain analyses, was efficiently reduced to 46 features, mitigating the curse of dimensionality while preserving essential information. As a result, the proposed method achieved a classification accuracy of over 96 %, with precision, recall, and F1-score of 0.95. • Comprehensive feature engineering for human arm movement using time and frequency-domain data from wearable IMU sensors. • Integrated dimensionality reduction (PCA, t-SNE) and embedded feature selection (Lasso, Random Forest, Elastic Net) pipeline. • Lasso and embedded methods selected 46 key features out of 720 , achieving over 96 % classification accuracy. • Gyroscopic data identified as critical for accurate arm movement classification. • Applications in rehabilitation , sports science, and real-time Range of Motion detection using wearable devices.