Accelerometer-Based Multi-Domain Feature Extraction for Parkinsonian Activity Recognition

This paper details our methodology for the ABC Conference 2025 challenge on human activity recognition using wearable accelerometer data. Research also explores the use of broader wearable data for monitoring and predicting Parkinson’s disease-related phenomena such as wearing-off [1]. The approach combines comprehensive feature extraction from diverse domains (statistical, energy, frequency, nonlinear and temporal) with Recursive Feature Elimination using Cross-Validation (RFECV) to select a highly relevant feature subset. The selected features are then used to train multiple classifiers, including XGBoost, Random Forest, Multi-Layer Perceptron (MLP), and Self-Training Random Forest. Among these, XGBoost achieved the best performance and was ultimately used to classify ten different activities, including Parkinson’s-related movements. The evaluation results indicate that the XGBoost model achieves a high overall accuracy and a weighted F1-score of $40 \%$ on the challenge dataset. The contribution of key features, such as spectral centroid and Petrosian fractal dimension, to classification performance is discussed. Limitations of the current methodology and potential future work, including the incorporation of deep learning techniques and multi-modal data integration to enhance robustness and generalizability for Parkinson’s activity monitoring, are also addressed.