Activity Recognition in Older Adults with Training Data from Younger Adults: Preliminary Results on in Vivo Smartwatch Sensor Data
Sabahat Fatima · 2021 · Proceedings of the 23rd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS '21) · doi:10.1145/3441852.3476475
Summary
This extended abstract investigates a critical age-related bias in wearable activity recognition: models trained on data from younger adults perform significantly worse when applied to older adults. The study is motivated by the growing potential of smartwatch-based self-tracking to help older adults reduce sedentary behaviours and engage in physical activity — important for preventing chronic diseases like diabetes and heart disease. However, the three main smartwatch human activity recognition (HAR) datasets (WISDM, UCI-HHAR, and Extrasensory) represent participants aged 18-48, with none including older adults over 65. The author developed a hybrid LSTM-CNN (Long Short-Term Memory and Convolutional Neural Network) model trained on the Extrasensory dataset, which contains data from 60 users aged 18-42 with 300,000 samples across 51 activity labels collected in real-world settings. The model was designed to classify five primary activities: Sitting, Standing, Walking, Running, and In-Vehicle. To evaluate performance on older adults, the researcher collected week-long in-the-wild smartwatch accelerometer data from two older adult participants (PP1, PP2), using the activPAL thigh-worn sensor as ground truth. The study also compared results against the Google Activity Recognition API and a Multilayer Perceptron baseline model.
Key findings
The results confirmed the hypothesis that models trained on younger adult data generalise poorly to older adults. The LSTM-CNN model showed approximately 25% greater accuracy and 11% greater balanced accuracy on younger participants compared to older participants across the five activity labels. The performance gap was particularly stark for the Standing activity, which had a poor 51% average balanced accuracy for older adults. Walking and Running showed the largest accuracy drops, consistent with prior literature showing fitness trackers inaccurately reporting older adult activities performed at slower speeds. Compared to the Google Activity Recognition API using activPAL ground truth, the custom LSTM-CNN model performed comparably or better for Walking (+4% accuracy, +1% balanced accuracy), Running (+4%, +1%), and Standing (+46%, +12%). However, the Google API outperformed for In-Vehicle (by 2.5% accuracy) and Sitting (by 5% balanced accuracy) — two activities that depend more on lower body motion, where the thigh-worn activPAL ground truth provides better signal than the wrist-worn smartwatch. The author acknowledges limitations including only two older adult participants, imbalanced activity data, and differences in ground truth methods between age groups.
Relevance
This research highlights an important and often overlooked form of bias in accessible technology: age-based dataset bias in wearable activity recognition systems. As smartwatches and fitness trackers become increasingly important health monitoring tools for older adults, the finding that these systems are fundamentally trained on younger bodies has direct implications for their reliability and usefulness. The 25% accuracy gap means that older adults — the very population that could benefit most from activity tracking to combat sedentary behaviour — receive the least accurate service. The paper advocates for building training datasets that include older adult representation and explores personalisation approaches such as transfer learning, meta-learning, and teachable machines that could allow models to adapt to individual movement characteristics. For accessibility practitioners, this work serves as a reminder that age-related differences in movement patterns constitute a form of diversity that must be represented in training data, paralleling similar concerns about disability representation in AI datasets.
Tags: older adults · activity recognition · machine learning · wearable technology · smartwatch · self-tracking · aging · sedentary behaviour · dataset bias