Chen Huihe, Ling Tongsheng, Huang Lanhui, Wang Ling, Guan Xuehai, Gao Ming, Wang Zhao, Lan Wei, Xu Jian-Wen, Wei Zhuxin
Department of Emergency, Wuming Hospital of Guangxi Medical University, Nanning, Guangxi Province, China.
School of Computer Electronical and Information, Guangxi University, No.100, East Daxue Road, Xixiangtang District, Nanning, Guangxi Province, China.
BMC Geriatr. 2025 Sep 26;25(1):724. doi: 10.1186/s12877-025-06371-0.
Models that detect fall risk have been proposed. However, the value of an indicator derived from such models in fall-severity stratification is understudied. This study developed a machine learning (ML)-based fall classification model, constructed a fall-risk score, and explored its association with fall-related adverse outcomes.
We used the eXtreme Gradient Boosting algorithm to build a fall classification model using data from 15,457 community-dwelling adults aged 60 Years and older. Of the 216 fall-associated variables, the 15 most important variables were selected for modelling, and their directional relationships with falls were evaluated using the SHapley Additive exPlanation (SHAP) value. An ML-based fall-risk score (ML-FRS) was generated. Multilevel regression analysis was used to measure the associations between the ML-FRS and fall-related adverse outcomes, defined as recurrent falls or falls requiring treatment, in a subset of 3,514 participants.
Participants had a mean age of 85.4 Years, with 56.3% being women, and a 22.5% prevalence of a fall history. Women and older participants were more Likely to fall and experience fall-related adverse outcomes. Inability to stand up from sitting in a chair was the most important predictor of increased fall risk. A small calf circumference and a low plant-based diet score were associated with increased fall risk. The ML-based model had an area under the curve of 0.797. Compared with non-fallers, participants in the highest ML-FRS quartile had a significantly higher risk of one fall without treatment, recurrent falls without treatment, one fall with treatment, and recurrent falls with treatment.
The ML-FRS could be used to screen for fall risk and fall-related adverse outcomes in community-dwelling older adults.
已经提出了检测跌倒风险的模型。然而，此类模型得出的指标在跌倒严重程度分层中的价值尚未得到充分研究。本研究开发了一种基于机器学习（ML）的跌倒分类模型，构建了跌倒风险评分，并探讨了其与跌倒相关不良结局的关联。
我们使用极端梯度提升算法，利用15457名60岁及以上社区居住成年人的数据构建跌倒分类模型。在216个与跌倒相关的变量中，选择15个最重要的变量进行建模，并使用夏普利值（SHAP）评估它们与跌倒的方向性关联。生成了基于ML的跌倒风险评分（ML-FRS）。在3514名参与者的子集中，采用多水平回归分析来衡量ML-FRS与跌倒相关不良结局（定义为反复跌倒或需要治疗的跌倒）之间的关联。
参与者的平均年龄为85.4岁，女性占56.3%，跌倒史患病率为22.5%。女性和老年参与者更有可能跌倒并经历与跌倒相关的不良结局。无法从椅子上坐立起来是跌倒风险增加的最重要预测因素。小腿围较小和植物性饮食得分较低与跌倒风险增加有关。基于ML的模型曲线下面积为0.797。与未跌倒者相比，ML-FRS四分位数最高的参与者未经治疗发生一次跌倒、未经治疗反复跌倒、经治疗发生一次跌倒和经治疗反复跌倒的风险显著更高。
ML-FRS可用于筛查社区居住老年人的跌倒风险和跌倒相关不良结局。
