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

Shuhao Que, Dieuwke van Dartel, Ilse Heeringa, Han Hegeman, Miriam Vollenbroek-Hutten, Ying Wang · 2026

Read Synthetic Data–Guided Feature Selection for Robust Activity Recognition in Older Adults by Shuhao Que, Dieuwke van Dartel, Ilse Heeringa, Han Hegeman, Miriam Vollenbroek-Hutten, Ying Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-17053.txt.

Introduction

Shuhao Que and a team of five co-authors keep gait tools true for old adults in rehab. Hip breaks need slow, tracked work to heal well. Wear tools can count steps and pace in that work. Most such tools were tuned on mid-age gaits. Old gaits are slower and less even, so tools slip. Their 2026 preprint uses fake data to pick sound traits.

Fake gait data fills gaps where real data is thin. Trait picks guided by fakes hold up on real old gaits. True counts help docs tune rehab with care. They also flag when a patient's pace slides back.

Catalog record

CREATORShuhao Que, Dieuwke van Dartel, Ilse Heeringa, Han Hegeman, Miriam Vollenbroek-Hutten, Ying Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-17053

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Synthetic Data–Guided Feature Selection for Robust Activity Recognition in Older Adults Shuhao Que, Dieuwke van Dartel, Ilse Heeringa, Han Hegeman, Miriam Vollenbroek-Hutten, Ying Wang Abstract 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 ≥ \geq 80 years were included as an age-matched feasibility cohort. Participants performed activities of daily living under simulated free-living conditions for 75 minutes while wearing two accelerometers positioned on the lower back and anterior upper thigh. To address and enhance model robustness to inter-person variability, a novel synthetic data generation scheme was proposed to capture common gait characteristics across individuals. Our real data and synthetic data have been published along with this work. Leveraging this synthetic dataset, a feature selection pipeline was subsequently applied to identify a compact set of six discriminative features for classifying walking, standing, sitting, lying down, and postural transfers. 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 ± \pm 0.100 for walking, 0.927 ± \pm 0.039 for standing, 0.997 ± \pm 0.004 for sitting, 0.937 ± \pm 0.202 for lying down, and 0.816 ± \pm 0.120 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. This paper has been submitted to Nordic-Digihealth 2026 conference for review. 1 Introduction Hip fractures pose a growing burden on healthcare systems due to population ageing [ 58 ] . Although surgical treatment is often successful, more than half of patients fail to regain prefracture mobility within one year [ 15 , 97 ] , resulting in reduced independence in activities of daily living (ADL) [ 71 ] and long-term declines in health-related quality of life (HRQoL) [ 46 ] . Increasing physical activity during rehabilitation is one of the few modifiable strategies shown to improve mobility and ADL independence [ 91 , 42 , 31 , 99 ] , and to reduce secondary fracture risk by preventing sarcopenia and balance impairments [ 92 , 30 , 44 , 40 ] . Despite this clinical importance, physical activity is rarely quantified in routine care [ 90 ] . Wearable activity trackers have therefore been introduced in postoperative hip fracture rehabilitation [ 78 , 43 ] , yet their accuracy remains uncertain, as they perform poorly for low-intensity movements such as slow or shuffling gait, which are prevalent in geriatric patients [ 66 , 43 ] . Recent validation work reinforces these concerns. In our previous study [ 96 ] , a commercial human-activity-recognition (HAR) algorithm [ 3 ] substantially underestimated walking time in hip-fracture patients aged 82 \pm 6 years, frequently misclassifying walking as standing. Such errors arise because most HAR algorithms are calibrated on middle-aged adults, whereas older adults present slower gait and altered movement amplitudes [ 8 , 101 , 54 ] . As a result, HAR models developed in younger populations often fail to capture ambulation behavior in geriatric users [ 34 ] . Accurate monitoring also requires recognizing the functional milestones that determine clinical discharge readiness, including sitting, standing, lying down, walking, and transfers (sit-to-stand, stand-to-sit, sit-to-lie, lie-to-sit) [ 62 ] . Yet most HAR studies involving older adults overlook transfers [ 32 , 70 , 28 , 86 , 8 ] . The only system addressing all milestones, developed by Allen et al. [ 2 ] , was not evaluated in free-living conditions (FLC), where individual habits and environmental unpredictability introduce substantial intra-activity variation [ 89 ] . Collecting data under FLC introduces its own challenges. Because participants perform activities at self-selected frequency and duration, sensor data become highly imbalanced [ 21 ] . Imbalances can limit the representation of diverse gait patterns and reduce robustness to inter-person variability—one of the major determinants of generalization failure in HAR [ 33 , 76 , 1 , 39 ] . This indicates the following methodological gap: HAR development for geriatric monitoring must account simultaneously for free-living variability and inter-subject gait differences. To address these gaps, we developed a new HAR algorithm for detecting hospital-discharge-relevant activities in the context of geriatric hip fracture rehabilitation. Specifically, we developed the HAR system on an age-appropriate cohort of healthy older adults aged \geq 80 years, whose gait characteristics better reflect the target clinical population [ 52 ] . Our study design captured intra-activity variation by collecting wearable-sensor data under simulated FLC rather than scripted laboratory tasks. To improve robustness to inter-person gait variability and reduce overfitting to idiosyncratic patterns under realistic free-living class imbalance, we proposed a novel synthetic data generation scheme, which captured common movement characteristics across individuals. Rather than explicitly rebalancing activity classes, class-specific synthetic data were generated to enrich the representation of shared gait patterns within each activity, thereby increasing the effective diversity of underrepresented classes and mitigating class-imbalance-induced overfitting in the feature selection process. The synthetic data generation scheme was developed to support more reliable physical activity monitoring during hip fracture rehabilitation in the future. 2 Methods 2.1 Data collection This prospective observational study was conducted at the eHealth House (eHH) of the TechMed Simulation Centre, University of Twente (Enschede, The Netherlands). The dataset is available upon request via Zenodo (restricted access, DOI: 10.5281/zenodo.18222937). The eHH is a controlled environment designed to simulate free-living conditions (FLC) and includes a living room, kitchen, bedroom, and bathroom. All experimental sessions were recorded using five fixed cameras installed throughout the eHH. Participants were eligible if they were aged 80 years or older and physically able to participate independently. Individuals with cognitive impairments or mobility disorders that impeded task performance were excluded. According to Dutch law, and based on a ruling from the Medical Research Ethics Committee (MREC) Arnhem–Nijmegen, the study was exempt from the Medical Research Involving Human Subjects Act. Ethical approval was granted by the Ethics Committee of the Natural Sciences and Engineering Sciences of the University of Twente. All participants provided written informed consent prior to participation. Each participant visited the eHH once for a 75-minute session. During the visit, participants performed a set of activities of daily living (ADL) at their own pace and in their preferred order. These tasks reflected functional discharge criteria for geriatric hip fracture patients and included: (1) walking inside the eHH (between the living room, kitchen, bedroom, and front door), (2) visiting the toilet once, (3) getting in and out of bed once

Read the complete text