✕
国家预印本平台基于多算法融合的跌倒检测模型
A Fall Detection Model Based on Multi-Algorithm Fusion
摘要
摘要：
[目的] 跌倒事件是无陪护病房患者监护的核心痛点，现有技术普遍存在误报率高、敏感性不足等局限。本研究旨在基于可穿戴设备构建高精度、低延迟的融合跌倒检测模型。
[方法] 采集并构建了跌倒数据集，通过中值滤波与动态窗口预处理后，提出“阈值初筛—随机森林（RF）静态判别—双向长短期记忆网络（Bi-LSTM）动态判别”的多算法融合框架。
[结果] 5折交叉验证结果显示，模型在独立测试集上的综合准确率达98.52%，精确率94.40%，召回率97.61%，F1-score 95.96%，系统延迟控制在1秒以内。
[局限] 模型泛化能力一定程度上受限于当前样本规模，在面对极度缓慢的侧滑等罕见非典型动作时仍有偶发误判，且尚未融合跨模态生理信息。
[结论] 该融合框架有效解决了穿戴式跌倒检测中计算开销与精度的矛盾，具备在边缘设备低功耗部署的潜力，具有极高的临床应用价值。
Abstract
[Objective] .Fall events are a critical pain point in patient monitoring in unescorted wards. Existing technologies often suffer from high false alarm rates and insufficient sensitivity. This study aims to build a high-precision, low-latency fusion fall detection model based on wearable devices. [Methods] A real-world dataset (FallData) with 9,611 samples was constructed using IM900 sensors. After preprocessing with median filtering and dynamic windows, a multi-algorithm fusion framework was proposed: "Threshold pre-screening — Random Forest (RF) static discrimination — Bidirectional Long Short-Term Memory (Bi-LSTM) dynamic modeling." [Results] The results of 5-fold cross-validation showed that the model achieved an overall accuracy of 98.52%, a precision of 94.40%, a recall rate of 97.61%, and an F1-score of 95.96% on the independent test set, with system latency controlled within 1 second. [Limitations] The model's generalization is somewhat constrained by sample size. Occasional misjudgments occur with atypical actions like slow sliding, and cross-modal information is not yet integrated. [Conclusions] This framework successfully resolves the conflict between computational overhead and accuracy in wearable fall detection, demonstrating great potential for low-power edge deployment and high clinical value.关键词
跌倒检测 可穿戴设备 随机森林 长短期记忆网络 智慧医疗Key words
Fall Detection/ Wearable Devices/ Random Forest/ Long Short-Term Memory Network/ Smart Healthcare引用本文复制引用
李俊儒,陆衢航,陈士凯,王京华.基于多算法融合的跌倒检测模型[EB/OL].(2026-04-15)[2026-04-16].https://sinoxiv.napstic.cn/article/25763079.学科分类
机电一体化基金
长春理工大学大学生创新训练计划 慧护安康——基于OpenHarmony的多模态健康监测手环系统研发( 202510186058 )
评论