Age Matters: Evaluating Cross Population Generalization in Wearable Sensor Based Human Activity Recognition

Raza M. O., Babar A., Elkiran H., Raheed J.

13th International Conference on Electrical and Electronics Engineering, ICEEE 2026, Antalya, Türkiye, 27 - 29 Nisan 2026, ss.596-603, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/iceee69936.2026.11598238
  • Basıldığı Şehir: Antalya
  • Basıldığı Ülke: Türkiye
  • Sayfa Sayıları: ss.596-603
  • Anahtar Kelimeler: Activity of Daily Living, Human Activity Recognition, Machine Learning, Sensor Data
  • İstanbul Medipol Üniversitesi Adresli: Evet

Özet

Human Activity Recognition (HAR) using wearable sensors has become an important component of many healthcare and assistive technologies. Most HAR models are trained on datasets collected from a specific age group (Adults, children, elderly people, etc.), which limits the ability of the HAR system to generalize across diverse populations. In this study, we investigate the impact of age differences on HAR model performance. We are using two publicly available accelerometer datasets representing younger adults and older adults. Firstly, we extracted rolling window statistics, and intensity feature Signal Magnitude Vector (SMV) from raw sensor data, then we analyzed movement patterns across both groups using the SMV. We also trained and evaluated 5 Machine learning models, namely Random Forest, Decision Tree, Multilayer Perceptron, and Gradient Boosting, on the same and cross-dataset configurations. We also trained deep learning algorithms namely: 1 Dimensional Convolutional Neural Network (1D-CNN) and Long Short-Term Memory (LSTM). Based on results, models trained and evaluated on the same population achieve high performance, with Random Forest reaching accuracies of 92.2% and 92.8% for the older and younger datasets, respectively. However, when models trained on one population are evaluated on the other, performance drops significantly. Results improved when the model was trained and tested on a combined dataset with age-awareness, achieving an accuracy of 91.8%. We also performed the feature abalation study and feature extracted from the raw sensor data can improve the model performance over only raw sensor data. These results show the importance of age-aware modeling in HAR systems.