An Ensembled Penalized Federated Learning Framework for Falling People Detection
Sizhe Rao, Runqiu Zhang, Sajal Saha, and Liang Chen · 2025
Read An Ensembled Penalized Federated Learning Framework for Falling People Detection by Sizhe Rao, Runqiu Zhang, Sajal Saha, and Liang Chen (2025) 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-2510-20960.txt.
Introduction
Sizhe Rao leads three co-authors on EPFL, a federated learning framework for detecting falls in elderly and disabled people from wearable sensors. Falls cause heavy injury and death worldwide, and detection must stay accurate while keeping personal motion data private.
Centralized systems struggle with privacy, weak generalization, and wide variation in how people move. EPFL joins continual learning, personalized models, and a specialized weighted aggregation rule so sites learn together without pooling raw data. Sequential motion patterns are captured on device, aiming at fall alerts that adapt to each wearer while respecting data limits.
Catalog record
| CREATOR | Sizhe Rao, Runqiu Zhang, Sajal Saha, and Liang Chen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-20960
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Ensembled Penalized Federated Learning Framework for Falling People Detection Sizhe Rao, Runqiu Zhang, Sajal Saha, and Liang Chen Abstract Falls among elderly and disabled individuals remain a leading cause of injury and mortality worldwide, necessitating robust, accurate, and privacy-aware fall detection systems. Traditional fall detection approaches, whether centralized or point-wise, often struggle with key challenges such as limited generalizability, data privacy concerns, and variability in individual movement behaviors. To address these limitations, we propose EPFL—an Ensembled Penalized Federated Learning framework that integrates continual learning, personalized modeling, and a novel Specialized Weighted Aggregation (SWA) strategy. EPFL leverages wearable sensor data to capture sequential motion patterns while preserving user privacy through homomorphic encryption and federated training. Unlike existing federated models, EPFL incorporates both penalized local training and ensemble-based inference to improve inter-client consistency and adaptability to behavioral differences. Extensive experiments on a benchmark fall detection dataset demonstrate the effectiveness of our approach, achieving a Recall of 88.31% and an F1-score of 89.94%, significantly outperforming both centralized and baseline models. This work presents a scalable, secure, and accurate solution for real-world fall detection in healthcare settings, with strong potential for continuous improvement via its adaptive feedback mechanism. I Introduction Due to changes in traditional family structures, the number of older individuals living alone has significantly increased over the past few decades [ 1 ] . According to the report from World Health Organization (WHO) [ 2 ] , falls are the second leading cause of unintentional injury deaths worldwide, with particularly high morbidity among individuals aged 60 and older. Resulting in severe injuries, including fractures, head trauma, and even death, falls can significantly decline the quality of life of older adults [ 3 ] . Considering this, the need for effective monitoring and fall detection systems has been raised by this change aiming to ensure the safety of seniors. Falls can have long-term impacts on individuals, including significant disability-adjusted life years (DALYs) and high financial costs. According to the report [ 2 ] , falls cause over 38 million DALYs lost annually worldwide. In Canada, a 20% reduction in falls could save approximately US$120 million each year. Considering the severe injuries, potential fatalities and other additional costs resulting from sudden falls [ 4 ] , fall detection is a critical research area, especially for the elderly and individuals with disabilities. Effective fall detection methods are essential to reduce the severity and risk of falls. The reduction could essentially boost the quality of life for individuals by ensuring their safety and maintaining their functional independence. In practical circumstances, wearable devices (such as smart watches and fitness trackers) and ambient sensors placed in the living environment could help to monitor and detect sudden changes in movement and orientation. By utilizing the sequential position data from wearable devices and video/image data from ambient sensors, machine learning models can efficiently detect falls by treating the problem as time-series anomaly detection and human motion tracking respectively. In this paper, we will mainly focus on sequential-based fall detection given the data collected from wearable devices. Currently, there are several rough challenges in fall detection: • Accuracy and Reliability: Achieving high accuracy remains a major challenge in fall detection research. In some scenarios, it is hard to differentiate between normal activities and actual falls, such as lying down quickly and sudden falls [ 5 ] . The mis-recognition could lead to high false-positive or false-negative rates, which undermines the reliability of the detection system. Considering above, we would need to find the trade-off to balance the accuracy and reliability [ 6 ] to ensure the effectiveness of fall detection systems. • Privacy Concerns: Fall detection systems often rely on various sensors, including wearable devices and ambient cameras, to monitor individual movements and detect falls. However, these systems raise significant privacy and data security concerns due to the collection, storage, and processing of sensitive personal data [ 7 ] . In addition, the risk of data breaches or unauthorized access to sensitive information further heightens these concerns [ 8 ] . Therefore, novel approaches are required to respect user privacy while ensuring the effectiveness of these systems. • Individual Behaviour Differences: Variability in movement patterns, physical abilities, and daily routines among users can significantly affect the accuracy of fall detection. These behavioural differences can lead to reliability reduction in fall detection systems. To address this problem, personalized fall detection system [ 9 ] could probably help to more accuractely detect falls across diverse individuals to ensure higher reliability and effectiveness in real-world fall detection applications. Many studies have tackled these issues. [ 10 ] applied Multi-step Histogram-based Outlier Scores to achieve 12.85% Recall and 18.73% F1-score while [ 31 ] employed Incremental Kolmogorov-Smirnov Test to reach around 50% Accuracy. [ 12 ] attained 72% Accuracy with machine learning agents. Using the same dataset we used as our project, these studies indicate room for optimization. [ 13 , 14 ] addressed privacy concerns using federated learning; however, their approaches require client-side model training and fail to account for individual behavioral differences. To further address above problems, we propose a novel approach for falling people detection and the main contributions are stated below: • We propose EPFL—an Ensembled Penalized Federated Learning framework for fall detection, designed to address the challenges of privacy preservation, model robustness, and user-specific variability. EPFL integrates homomorphic encryption, penalized local training via FedProx, and ensemble-based inference to improve detection performance in real-world environments. • We develop a novel Specialized Weighted Aggregation (SWA) strategy that combines trimmed mean filtering, training-epoch normalization, and exponential moving average fusion. This improves the robustness and stability of the federated learning process in the presence of noisy or non-IID client data. • We introduce a dual-model ensemble approach that averages predictions from both global and client-specific models during inference. This personalization strategy significantly improves consistency and recall across heterogeneous clients without compromising privacy. • EPFL incorporates a user feedback-driven continual learning loop, allowing both global and local models to evolve over time based on new user-verified events. This ensures adaptive and responsive performance in dynamic healthcare settings. • Through extensive experiments on a publicly available fall detection dataset, our approach demonstrates superior performance (Recall = 88.31%, F1-score = 89.94%) compared to both classical models and existing federated learning baselines, validating its effectiveness and real-world applicability. The remainder of this paper is structured as follows. Section II provides a literature review, discussing background knowledge related to fall detection systems and recent advancements using wearable sensors, ambient sensors, and federated learning techniques. Section III introduces the proposed Ensembled Penalized Federated Learning (EPFL) framework, including its architectural components, data preprocessing strategies, specialized aggregation method, and local inference process. Section IV describes the experimental setup, including dataset details, evaluation metrics, software and hardware environment, and presents a comprehensive performance analysis of both baseline and proposed models. Section V concludes the paper by summarizing the key findings, addressing limitations, and outlining potential directions for future work. II Literature Review Fall detection systems can be divided into two main parts: the client part and the server part. The client part represents the user devices for real-time monitoring and alerting, such as wearable sensors and ambient sensors. The server part handles