Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Device Technologies and Biomedical Robotics

Device Technologies and Biomedical Robotics - Poster Session C & D

W38 - Advancing Fall Detection in Older Adults Through Video-Based Deep Learning

Friday, October 10, 2025

10:00 AM - 11:00 AM PDT

Location: Exhibit Hall F, G & H

Presenting Author(s)

Ethan Henley

Machine Leaning Engineer
Virginia Tech Helmet Lab
Pittsburgh, Pennsylvania, United States

Co-Author(s)

Nicole E.-P Stark, PhD

Research Assistant Professor
Virginia Tech
Christiansburg, Virginia, United States
Ethan Henley

Machine Leaning Engineer
Virginia Tech Helmet Lab
Pittsburgh, Pennsylvania, United States
JN

John Nowinski

Test Engineer
Virginia Tech Helmet Lab, United States
Brianna Reilly

Test Engineer
Virginia Tech Helmet Lab, United States
KM

Katherine D. Mayes, MD, PhD

Assistant Professor
Virginia Tech Carilion School of Medicine, Emergency Medicine
Roanoke, Virginia, United States
Damon R. Kuehl, MD

Professor, Emergency Medicine
Virginia Tech Carilion School of Medicine
Roanoke, Virginia, United States

Last Author(s)

SR

Steve Rowson

Associate Professor
Virginia Tech, United States

Introduction: : Older adults (65+ years) experience the highest rates of fall-related emergency department (ED) visits, hospitalizations, and fatalities in the United States[1-4]. As this population is projected to reach 83.7 million older adults by 2050 in the U.S., falls represent a growing public health crisis, burdening patients, caregivers, and the healthcare systems alike[5,6]. Approximately one in four older adults experiences a fall each year, totaling an estimated 14 million incidents. These events result in over 3 million ED visits, 1 million hospitalizations, and more than 36,000 deaths annually, with total medical costs exceeding $50 billion [5]. Beyond financial impact, falls can lead to serious health consequences, including fractures, traumatic brain injury, functional decline, and increased mortality.

Identifying falls in older adults presents a critical bioengineering challenge, as approximately 50% of falls in this population go unreported to medical providers[7]. This underreporting poses a significant barrier to timely intervention and secondary prevention. Traditional fall detection methods, such as patient or witness self-report and wearable alert systems, have critical limitations. Self-reports are hindered by cognitive impairments, fear of loss of independence, and/or recall bias[8]. Wearable devices suffer from poor adherence and limited real-world effectiveness[9].

Video-based monitoring has emerged as a promising alternative, offering continuous observation without requiring user engagement. Computer vision systems can detect and classify cues using spatiotemporal visual data, potentially overcoming key limitations of existing approaches. We aimed to develop and validate a deep learning-based computer vision monitoring system to detect falls that integrates local motion and contextual information.

Materials and

Methods: : We developed an innovative dual-stream architecture integrating a 3D Convolutional Neural Network (3DCNN) with a Vision Transformer (ViT) to enhance fall detection performance through simultaneous modeling of localized motion and global scene context. This approach addresses the inherent complexity of distinguising true falls from non-fall activities in older adults.

Training was conducted using two established datasets: the Le2i Fall Detection dataset [10] and a curated set of long-term care fall surveillance videos [11]. All videos were segmented into 5-second clips and labeled as fall or non-fall events. The training dataset consisted of 325 videos (203 fall and 203 non-fall videos for a total of 406 samples), augmented to 15,600 clips using a multi-dimensional augmentation strategy, to enhance model robustness and generalizability. This included spatial transformations (e.g. cropping, flipping), lighting and background variation (to simulate environmental diversity), motion blur, temporal jittering, and randomized frame dropout to mimic real-world video degradation. Our model architecture also implemented a learning strategy specifically engineered to address the challenges of fall detection with limited training data. The approach integrates three complementary loss components, each targeting distinct aspects of the learning process.

Model performance was evaluated using standard binary classification metrics: accuracy, precision, sensitivity, F1 score, and the area under the receiver operating characteristic curve (AUC-ROC). A confusion matrix was generated to visualize the distribution of true positives, true negatives, false positives, and false negatives.

Results, Conclusions, and Discussions:: The primary challenge in automated fall detection lies in distinguishing sudden, abnormal movements (e.g. falls) from routine or ambiguous actions (e.g. sitting abruptly, bending). Our model addressed this by combining the fine-grained spatiotemporal capabilities of 3DCNNs with the global reasoning power of ViT. This hybrid architecture successfully leveraged local motion dynamics while incorporating contextual cues to differentiate true fall events from non-falls.

Our system achieved robust performance: AUC-ROC = 0.92, F1 score = 0.89, precision = 0.88, recall = 0.90, and specificity = 0.94. The confusion matrix (Figure 1) revealed a true positive rate of 47.1% and a true negative rate of 45.1%, with false positive and false negative rates of 4.8% and 3.0%, respectively. These results demonstrate strong overall classification performance, particularly in minimizing false alarms, crucial for real-world deployment in clinical or home care environments.

Video-based monitoring systems represent a confluence of computer vision, biomechanical analysis, and clinical insight, offering a scalable and unobtrusive approach to fall surveillance. However, limitations remain. Most notably, our training dataset lacks footage from private, in-room environments, the most common setting for falls among older adults in long-term care. Incorporating these real-world contexts is critical for improving generalizability.

Additionally, detailed analysis of misclassified events is necessary to identify failure points—such as lighting, occlusions, or atypical fall postures—that challenge model robustness. Future iterations should incorporate domain-specific augmentation and architectural refinement to address these gaps.

In conclusion, this work demonstrates the feasibility of using a dual-stream computer vision model for passive fall detection in older adults. By enabling accurate identification of falls, such systems can support timely clinical intervention, reducing recurrent fall risk, and help interrupt the cascade of function decline associated with fall-related injuries. This work serves as a foundation for further development of intelligent video-based monitoring solutions aimed at improving safety and outcomes for older adults.

Acknowledgements and/or References (Optional): : References
[1] Houry, D., Florence, C., Baldwin, G., Stevens, J., & McClure, R. (2016). The CDC Injury Center’s Response to the Growing Public Health Problem of Falls Among Older Adults. American Journal of Lifestyle Medicine, 10(1), 74–77. https://doi.org/10.1177/1559827615600137
[2] Knowlton, L. M., & Staudenmayer, K. L. (2020). Traumatic Injury in Older Adults. In Principles and Practice of Geriatric Surgery: Third Edition: With 261 Figures and 155 Tables (pp. 277–297). Springer, Cham. https://doi.org/10.1007/978-3-319-47771-8_18
[3] Moreland, B., Kakara, R., & Henry, A. (2020). Trends in Nonfatal Falls and Fall-Related Injuries Among Adults Aged ≥65 Years — United States, 2012–2018. MMWR. Morbidity and Mortality Weekly Report, 69(27), 875–881. https://doi.org/10.15585/mmwr.mm6927a5
[4] Rockett, I. R. H., Regier, M. D., Kapusta, N. D., Coben, J. H., Miller, T. R., Hanzlick, R. L., … Smith, G. S. (2012). Leading causes of unintentional and intentional injury mortality: United States, 2000-2009. American Journal of Public Health, 102(11). https://doi.org/10.2105/AJPH.2012.300960
[5] Florence, C. S., Bergen, G., Atherly, A., Burns, E., Stevens, J., & Drake, C. (2018). Medical Costs of Fatal and Nonfatal Falls in Older Adults. Journal of the American Geriatrics Society, 66(4), 693–698. https://doi.org/10.1111/jgs.15304
[6] Stevens, J. A. (2002). Falls Among Older Adults: Public Health Impact and Prevention Strategies. Generations. American Society on Aging. Retrieved from http://www.safetylit.org/citations/index.php?fuseaction=citations.viewdetails&citationIds[]=citjournalarticle_514565_38
[7] Mckee, K. J., Orbell, S., Austin, C. A., Bettridge, R., Liddle, B. J., Morgan, K., & Radley, K. (2002). Fear of falling, falls efficacy, and health outcomes in older people following hip fracture. Disability and Rehabilitation, 24(6), 327–333. https://doi.org/10.1080/09638280110093686
[8] Ganz DA, Bao Y, Shekelle PG, Rubenstein LZ. Will My Patient Fall? JAMA. 2007;297(1):77–86. doi:10.1001/jama.297.1.77
[9] Klenk J, Schwickert L, Mackintosh SF, et al. Wearable sensor technology to detect falls in older adults: a systematic review and meta-analysis. PLoS One. 2016;11(4):e0154093. doi:10.1371/journal.pone.0154093
[10] Charfi, I., Miteran, J., Dubois, J., Atri, M., and Tourki, R., 2013, “Optimized Spatio-Temporal Descriptors for Real-Time Fall Detection: Comparison of Support Vector Machine and Adaboost-Based Classification,” J. Electron. Imaging, 22(4), p. 041106. https://doi.org/10.1117/1.JEI.22.4.041106.
[11] Robinovitch, S., 2018, “Falls Experienced by Older Adult Residents in Long-Term Care Homes.” https://doi.org/10.17910/B7.739.