Synthetic Data Guided Feature Selection for Robust Activity Recognition in Older Adults

Physical activity during hip fracture rehabilitation is essential for mitigating long-term functional decline in geriatric patients. However, it is rarely quantified in clinical practice. Existing continuous monitoring systems with commercially available wearable activity trackers are typically developed in middle-aged adults and therefore perform unreliably in older adults with slower and more variable gait patterns. This study aimed to develop a robust human activity recognition (HAR) system to improve continuous physical activity recognition in the context of hip fracture rehabilitation. 24 healthy older adults aged over 80 years were included to perform activities of daily living (walking, standing, sitting, lying down, and postural transfers) under simulated free-living conditions for 75 minutes while wearing two accelerometers positioned on the lower back and anterior upper thigh. Model robustness was evaluated using leave-one-subject-out cross-validation. The synthetic data demonstrated potential to improve generalization across participants. The resulting feature intervention model (FIM), aided by synthetic data guidance, achieved reliable activity recognition with mean F1-scores of 0.896 for walking, 0.927 for standing, 0.997 for sitting, 0.937 for lying down, and 0.816 for postural transfers. Compared with a control condition model without synthetic data, the FIM significantly improved the postural transfer detection, i.e., an activity class of high clinical relevance that is often overlooked in existing HAR literature. In conclusion, these preliminary results demonstrate the feasibility of robust activity recognition in older adults. Further validation in hip fracture patient populations is required to assess the clinical utility of the proposed monitoring system.

💡 Research Summary

The paper addresses a critical gap in continuous physical activity monitoring for older adults undergoing hip‑fracture rehabilitation. Existing wearable‑based human activity recognition (HAR) systems are typically trained on middle‑aged populations and therefore perform poorly when applied to seniors who exhibit slower, more variable gait patterns. To develop a robust HAR solution suitable for this clinical context, the authors recruited 24 healthy participants aged 80 years or older and recorded accelerometer data from two body locations: a thigh‑mounted MOX sensor (25 Hz) and a lower‑back APDM sensor (128 Hz). Data collection took place in a controlled “free‑living” environment (the eHealth House) that mimics a home setting, and activities of daily living—including walking, standing, sitting, lying, and postural transfers—were performed at self‑selected pace and order. Video recordings were annotated by two independent reviewers to provide ground‑truth labels, which were synchronized with the sensor streams.

Pre‑processing involved Savitzky‑Golay smoothing (0.12 s window, second‑order polynomial) to suppress high‑frequency noise while preserving signal peaks, followed by segmentation into non‑overlapping 2‑second windows. Windows containing mixed labels were discarded to avoid ambiguity. The core methodological innovation is the generation of class‑specific synthetic data using Dynamic Time Warping Barycentre Averaging (DBA). For each activity, at least 100 windows were gathered; one window per participant was randomly sampled (or, when insufficient data existed, a subset of six participants was sampled to increase combinatorial diversity). DBA was applied separately to each of the six sensor axes (thigh x, y, z; back x, y, z), aligning temporal variations and producing a representative “average” time series that captures common temporal patterns while suppressing individual idiosyncrasies. For walking, 4‑second windows were used to exploit periodicity, with the central 2 seconds retained; for transfers, variable‑length windows were normalized to a consistent length; static activities were directly processed as 2‑second windows. The synthetic dataset preserved the original class imbalance to ensure fair comparison with real data.

Feature selection was performed exclusively on the synthetic dataset, under the hypothesis that synthetic data would highlight subject‑invariant characteristics. An initial pool of over 150 time‑ and frequency‑domain features (mean, variance, energy, peak intervals, etc.) was reduced through a multistage pipeline: correlation‑based pruning, L1‑regularization, and recursive feature elimination, ultimately yielding a compact set of six discriminative features. These features were then extracted from the real recordings and used to train a classifier (random forest) within a Leave‑One‑Subject‑Out (LOSO) cross‑validation framework. The authors also built a control condition model (CCM) that followed the standard activity‑recognition chain (ARC) using only real data, with feature selection performed on a 25 % training split and model evaluation on the remaining 75 % of data.

Performance evaluation demonstrated that the synthetic‑data‑guided Feature Intervention Model (FIM) achieved high mean F1‑scores across all activities: walking 0.896 ± 0.100, standing 0.927 ± 0.039, sitting 0.997 ± 0.004, lying 0.937 ± 0.202, and postural transfers 0.816 ± 0.120. Notably, the detection of transfers—clinically crucial for discharge readiness—was significantly better than the CCM (p < 0.05). These results indicate that synthetic data effectively augment under‑represented classes, mitigate over‑fitting to participant‑specific patterns, and improve generalization to unseen subjects.

The study’s limitations include the modest sample size (24 participants) and the use of healthy older adults rather than actual hip‑fracture patients, which may limit direct clinical translatability. Additionally, DBA’s averaging nature could smooth out extreme gait abnormalities that are relevant in pathological populations. Future work is suggested to (1) incorporate real patient data, (2) compare DBA‑generated synthetic data with other generative approaches such as GANs or VAEs, (3) expand the activity taxonomy to finer‑grained transfer sub‑classes, and (4) integrate the HAR system into real‑time monitoring platforms for bedside decision support.

In conclusion, the authors provide compelling evidence that synthetic data‑guided feature selection can produce a robust, generalizable HAR system for older adults, especially improving the detection of postural transfers. This approach offers a promising pathway toward reliable, continuous activity monitoring in hip‑fracture rehabilitation, pending validation in clinical patient cohorts.