پیش بینی سقوط سالمندان با استفاده از تک حسگر اینرسی در حین راه رفتن : یک مرور روایتی

Introduction

Falls in the elderly population represent a critical public health challenge with severe consequences including injuries, loss of independence, and increased mortality. Falls are the second leading cause of accidental deaths and affect approximately 40% of elderly individuals in daily activities. In Canada, falls are the primary cause of injury in older adults, with 33% of hospitalized cases transferred to long-term care facilities. Risk factors encompass intrinsic elements (physiological and psychological) and extrinsic factors (environmental).

Traditional assessment tools such as the Timed Up and Go (TUG) test and Berg Balance Scale provide qualitative fall risk evaluation, whereas wearable sensors offer quantitative assessment. Inertial measurement units (IMUs) provide precise measurement of movement pattern changes and body posture during walking. Research activity has increased substantially since 2013, with significant advances in machine learning applications. Bibliometric analysis identified 221 studies in this field from 2000–2024, with the Journal of Neurology and IEEE Transactions publishing the most articles.

Methods

This narrative review analyzed evidence regarding inertial sensor applications in elderly fall risk assessment. A narrative approach was selected due to the heterogeneous nature of research in this emerging technology field, enabling deep qualitative analysis of diverse methodologies.

Structured searches were conducted in Web of Science and Scopus using keywords: (wearable sensor OR inertial measurement unit OR IMU OR accelerometer) AND (fall risk OR fall prediction OR fall detection) AND elderly. Time frame: 2015–2025; English language. Inclusion criteria: cross-sectional, case-control, prospective cohort, validation, and algorithm development studies using IMUs in elderly (≥60 years) reporting fall-related outcomes. Exclusion criteria: review studies, conference abstracts, non-wearable sensors, populations <60 years, simulations, and animal studies. Twelve studies met criteria and underwent detailed analysis.

Data extraction included: study characteristics, demographics, sensor specifications (type, placement location, sampling frequency), assessment protocols, extracted variables (stride length, gait speed, Lyapunov exponent, entropy), analytical methods, and outcomes (prediction accuracy, sensitivity/specificity).

Results

Sensor Placement

IMUs positioned in core body regions, particularly lower lumbar vertebrae (3rd–5th), demonstrated optimal effectiveness. Wang et al. (2024) achieved 88% area under receiver operating characteristic curve, the best discrimination between high and low-risk groups. Buisseret et al. (2022) achieved 76% accuracy with 4th lumbar vertebral placement. This location's superiority stems from high sensitivity to trunk changes and body stability during position transitions. Liu (2012) achieved 86.7% sensitivity with sternum placement versus 73% for foot-based sensors (Neira, 2023), due to time-series analysis independent of step detection, eliminating errors in shuffling gait patterns common in elderly. However, Saadeh et al. achieved 98% prediction accuracy with thigh sensors but had limited generalizability due to small sample size (20 subjects) and lack of prospective validation.

Assessment Protocols

The TUG test was used in 7 of 12 studies and revealed specific deficits when instrument-augmented. Qiu et al. integrated three assessment domains (TUG, stability limits test, five-time sit-to-stand) achieving 89.4% accuracy using Support Vector Machine, substantially outperforming simpler approaches using only gait speed analysis (Bautmans: 2011 variance analysis only).

Diverse assessment methods showed significant differences. Lockhart's 10-meter walk test with prospective design (6-month follow-up) achieved 81.6% predictive accuracy with better generalizability than Neira's 15-minute free walking (73% accuracy, case-control). Real-world monitoring provided greater validity: van Schooten et al. (2015) conducted one-week ambulatory monitoring using logistic regression, while Handelzalts et al. (2020) identified 18 of 22 trip events (82%) in daily-life conditions. In contrast, Rivolta et al. (2019) achieved 89% laboratory accuracy with artificial neural networks but lacked real-world applicability.

Gait Parameters and Feature Extraction

Linear spatiotemporal gait parameters consistently discriminated fallers from non-fallers: gait speed in 9 of 12 studies, variability indices in 6 of 12, and step frequency in 7 of 12. Liu's combined approach using linear features (step timing) with nonlinear indices (multiscale entropy and recurrence quantification analysis) achieved 81.6% prospective accuracy, outperforming purely linear approaches (Bautmans, van Schooten). Nonlinear indices demonstrate superior sensitivity to subtle gait deterioration. Howcroft achieved 84% accuracy using maximum Lyapunov exponent in a neural network. Qiu achieved 89.4% using frequency-domain features in Support Vector Machine frameworks.

However, linear variability measures depend on accurate step detection and fail in shuffling gait patterns. Frequency-domain features are environment-sensitive, as Wang demonstrated with superior discrimination in stair descent versus level walking.

Algorithm Comparison

Ensemble tree methods (Random Forest) and margin-based methods (Support Vector Machine) provide optimal balance between accuracy and generalizability. Rivolta's neural network was trained on Tinetti scores, an imperfect criterion not fully reflecting real-world fall occurrence.

Prospective vs. Retrospective Validation

Prospective studies demonstrated superior clinical prediction validity. Liu's Random Forest algorithm with 6-month prospective follow-up achieved 81.6% accuracy and proved more clinically reliable than Qiu's Support Vector Machine (89.4%) based on internal cross-validation without external testing. Neira's Support Vector Machine using foot sensors (73% accuracy) was limited by retrospective case-control design, reducing predictive power for future falls. Real-world monitoring advantages are evident: Handelzalts' ecological validity (82% sensitivity) versus simplified laboratory environments demonstrates superior clinical utility, though with limitations such as missed events (4 of 22 self-reported trips undetected).

Conclusion

IMUs positioned optimally in core body regions, particularly lower lumbar spine (3rd–5th vertebrae), represent effective fall risk assessment tools achieving 76–89.4% accuracy. Combined TUG and stability limit tests with linear (speed, stride length, variability) and nonlinear indices (multiscale entropy, Lyapunov exponent) provide superior discrimination. Machine learning algorithms including Support Vector Machine and Random Forest demonstrate higher performance than decision trees. Prospective studies (Liu: 81.6% over 6 months) show superior validity compared to internal cross-validation (Qiu: 89.4%), while real-world monitoring provides greater ecological validity.

Primary limitations include small sample sizes causing model overfitting, lack of standardization regarding sensor placement and parameters, and poor data quality (mostly simulated laboratory falls). Future directions include: establishing public databases with standardized protocols, integrating diverse sensors for comprehensive monitoring, utilizing Internet of Things with personalized feedback, and developing advanced analytical methods (deep reinforcement learning, recurrent neural networks) for detecting dynamic fall risk patterns in real-world environments.

Footnotes

Authors’ contribution

Study concept and design: E. SH.; Analysis and interpretation of data: Y. D..; Drafting of the manuscript: Y. D..; Critical revision of the manuscript: E. SH.

Funding

Non.

Conflict of interest

According to the authors, this article has no conflict of interest.