通讯作者: 高建强, jianqianggaohh@126.com; 李莉, liliiot@163.com

DOI：10.12201/bmr.202607.00038

Patient-Level Early Identification of Freezing of Gait Using a Streamlined Temporal Model Based on Inertial Measurement Unit Data

• 摘要：目的/意义 构建面向未见患者的可穿戴惯性测量单元冻结步态早期识别模型，评估其跨患者稳定性以及关键特征和网络模块的作用。方法/过程 基于12名患者的56条公开记录，采用传感器可用性核查和方向无关小腿特征汇聚，构建融合残差一维卷积、双向长短期记忆网络和注意力池化的精简惯性测量单元时序模型，并采用3个固定随机种子进行患者级留一内部验证和消融分析。结果/结论 模型的准确率、平衡准确率、F1值、AUROC和AUPRC分别为0.7910、0.7378、0.6325、0.8791和0.7595。冻结—步态频带特征及双向时序建模对模型性能贡献较大，在本数据集的患者级留一验证中表现出一定的跨患者稳定性。

  关键词： 帕金森病; 冻结步态; 惯性测量单元; 早期识别; 患者级交叉验证

   

  Abstract: Objective/Significance This study presents a wearable inertial measurement unit (IMU)-based framework for early freezing of gait (FOG) identification in previously unseen patients, specifically targeting cross-patient generalizability and dissecting the relative contributions of key spectral features and network modules. Methods/Process We interrogated a publicly available dataset comprising 56 recordings from 12 patients. After verifying sensor availability, we performed direction-invariant feature aggregation from shank-mounted IMU signals. A compact temporal architecture was subsequently designed, integrating residual one-dimensional convolutions, bidirectional long short-term memory (LSTM) networks, and an attention-based pooling mechanism. Internal validation employed a patient-level leave-one-subject-out (LOSO) scheme, alongside systematic ablation studies, all conducted across three fixed random seeds. Results/Conclusion The proposed pipeline yielded an accuracy of 0.7910, a balanced accuracy of 0.7378, an F1-score of 0.6325, an area under the receiver operating characteristic curve (AUROC) of 0.8791, and an area under the precision–recall curve (AUPRC) of 0.7595. The freeze- and gait-band spectral features, together with bidirectional temporal modeling, emerged as primary drivers of performance. Notably, across the three fixed seeds during LOSO internal validation on the current dataset, the AUROC and AUPRC exhibited only limited variability, confirming reliable cross-patient stability.

  Key words: Parkinson’s disease; freezing of gait; inertial measurement unit; early identification; patient-level cross-validation

  提交时间：2026-08-26

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 余啸；yu xiao;xiao yu;, 李达. 嗅觉测试联合多指标早期帕金森病诊断价值. 2025. doi: 10.12201/bmr.202512.00075

  李会君, 霍洁, 杭菲菲, 徐静, 管峥. 血清IgE、LDH及IL-18对儿童难治性肺炎支原体肺炎的早期识别价值. 2026. doi: 10.12201/bmr.202606.00009

  周栋, 朱思颖, 高欣怡, 孙禧斌, 袁蓓蓓. 基层卫生人员工作满意度与患者满意度的关系：卫生服务质量感知的跨层次中介作用. 2026. doi: 10.12201/bmr.202606.00082

  张敏, 吴胤歆. 我国医保基金协同监管的协同度测量与优化路径研究——基于SFIC模型. 2023. doi: 10.12201/bmr.202311.00006

  刘雨丹, 李慧芳, 李建春, 翟亚仙, 杨锦湄△, 沈云霞. 帕金森病患者嗅觉障碍影响因素的系统评价. 2025. doi: 10.12201/bmr.202503.00057

  梁乐然, 阎誉豪, 唐跃豪, 练洢彤, 徐倩, 杨猛, 伍家亨, 陈振虎, 刘秀峰. 基于知识图谱和大语言模型的中医药治疗帕金森病智能问答研究. 2026. doi: 10.12201/bmr.202604.00033

  韩乐阳, 刘晓云. 农村慢性病患者整合服务水平与基层就医选择的关联研究. 2026. doi: 10.12201/bmr.202608.00036

  张二利, 何兰兰, 李丹阳, 沈立, 吴钟华, 张军, 叶永强. 基于脑小血管病影像学总负荷的急性脑梗死患者住院时间延长列线图模型构建与验证. 2025. doi: 10.12201/bmr.202501.00008

  张宁, 玛依拉·阿不都克力木. AI工具在COPD患者继发肺动脉高压早期预测的应用. 2026. doi: 10.12201/bmr.202607.00013

  陈冰融, 舒文秀, 罗柳飞, 童嘉琦, 乐静. 骨髓瘤患者发生3-4级骨髓瘤骨病的风险预测模型构建. 2024. doi: 10.12201/bmr.202410.00057

• 序号	提交日期	编号	操作
  1	2026-07-24	10.12201/bmr.202607.00038V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论