運用機器學習建構社區老年人行動力衰退之預測模型

Developing a Predictive Model for Mobility Impairment in Community-Dwelling Older Adults Using Machine Learning

目的：老年人行動力受損與跌倒、衰弱及機構式照護等不良結局密切相關，及早發現高風險個案，及早介入，延緩失能為重要議題。台北市老人健檢是很普及的預防保健服務，民眾可了解自己的健康狀況，長期更可知道健康狀況變化。本研究目的為以機器學習方法建立老年人行動力受損風險的預測模型。

方法：本研究回溯分析2023年於臺北市立聯合醫院接受健檢與ICOPE行動力篩檢之2,165位65歲以上社區長者的健檢資料。經由ANOVA F值法選出13項具鑑別力的變項，涵蓋年齡、腎功能、血液指標、體位參數及慢性疾病，以Balanced Random Forest分類器進行機器學習模型訓練與五折交叉驗證。

結果：模型平均AUC-ROC為0.9671，PR-AUC為0.9350，精確度、召回率及F1分數分別為0.8549、0.8747及0.8646。行動力受損與年齡、肌酸酐、腰圍呈正相關，與血紅素及白蛋白呈負相關。

結論：本研究的成果可有效識別行動力受損高風險老年人，具備臨床可行性。此方法有助於將預測分析導入社區健康照護，早期發現功能下降及早介入以促進健康老化。

 

Purpose: Mobility impairment among elderly is closely associated with adverse outcomes such as falls, frailty, and institutionalization. Early identification of high-risk individuals and timely intervention are crucial for delaying the onset of disability. The Taipei City Elderly Health Examination is a widely utilized preventive health service that allows individuals to understand their current health status and monitor changes over time. The aim of our study is to develop a predictive model for the risk of mobility impairment in elderly using machine learning methods.

Methods: We conducted a retrospective analysis of 2,165 community-dwelling adults aged ≥65 years who underwent health check-ups and ICOPE mobility screening at Taipei City Hospital in 2023. Thirteen features were selected via ANOVA F-value, including age, renal and hematologic markers, anthropometrics, and chronic conditions. Machine learning model training was performed using the Balanced Random Forest classifier with five-fold cross-validation.

Results: The final model achieved a mean AUC-ROC of 0.9671 and mean PR-AUC of 0.935. Precision, recall, and F1-score were 0.8549, 0.8747, and 0.8646, respectively. Mobility impairment was positively associated with age, creatinine, and waist circumference, and negatively associated with hemoglobin, and albumin.

Conclusions: The proposed model of this study shows promise in identifying older adults at risk of mobility impairment. This approach supports the integration of predictive analytics into community health to enable timely, personalized interventions that promote healthy aging.