| 摘要: | 髖關節骨折好發於老年人身上，常見治療方式為手術治療，大多分為兩種，開放性復位和內固定與人工關節成形術（包括全關節成形術和半關節成形術），發生後12個月內的累計死亡率在20 %至40 %之間。透過收集大量臨床數據，我們的研究旨在利用年齡、性別、骨折型態、手術方式、術前抽血驗驗、其他内科病症等變數套用在機器學習上建立預測模型，藉以預測髖關節骨折手術後一年存活率，以更精確地評估老年病人在手術後的生存風險。這不僅有助於提前識別高風險病人，還可以為臨床醫生提供個別化的預防和治療建議。
我們使用了來自臺北醫學大學臨床研究資料庫(Taipei Medical University Clinical Research Database, TMUCRD)的資料，該資料包含臺灣北部地區三家醫院的病人資訊。研究中採用了隨機森林（Random Forest, RF）、支持向量機（Support Vector Machine, SVM）、邏輯斯回歸（Logistic Regression, LOG）、極端梯度提升（eXtreme Gradient Boosting, XGB）以及輕量梯度提升機（Light Gradient Boosting Machine, LGB）五種機器學習模型進行分析。
在顯著相關變項(來自於單便項羅吉斯回歸顯著因子)、中位數補值和LGB模型組合並且使用SMOTENC（Synthetic Minority Over-sampling Technique for Nominal and Continuous features）進行資料增量達到效能最佳，AUC達0.723。而SVM模型在任何情況下有較高的敏感度。
研究變項方面，我們採用了人口學和共病項目，與相關文獻一致。抽血檢驗資料的遺失值成為研究的挑戰之一，未來的研究可進一步探索更合適的處理方法。建議未來研究者在資料庫使用上進行更深入的培訓，考慮納入手術中的各項數值等因素，以提高模型的預測能力。
最後總結，雖然我們在資料庫使用上遇到一些困難，但本研究的結果有望為臨床實踐提供實用的參考，改善老年病人在髖關節置換手術後的生存率。同時，透過深入探討老年病人的手術後風險，我們可以更好地理解這一人群在醫學上的需求，為未來的醫療服務和政策制定提供有價值的參考。
Hip fractures are prevalent among the elderly and are commonly treated with surgical interventions, including open reduction and internal fixation or joint replacement surgeries. The cumulative mortality rate within 12 months post-surgery ranges from 20% to 40%. In this study, we aimed to utilize machine learning techniques on a large clinical dataset, including variables such as age, gender, fracture type, surgical approach, preoperative blood tests, and other comorbidities, to establish a predictive model for one-year survival after hip fracture surgery. This model aims to accurately assess the postoperative survival risk for elderly patients, facilitating early identification of high-risk individuals and providing personalized prevention and treatment recommendations for clinicians.
We utilized data from the Taipei Medical University Clinical Research Database (TMUCRD), which encompasses patient information from three hospitals in the northern region of Taiwan. Five machine learning models, namely Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LOG), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGB), were employed for analysis.
In our study, the optimal performance was achieved by combining significant variables identified through single-variable logistic regression, median imputation, and the (LGB) model. We utilized the Synthetic Minority Over-sampling Technique for Nominal and Continuous features (SMOTENC) to augment the data, resulting in an improved area under the curve (AUC) of 0.723. Notably, the Support Vector Machine (SVM) model exhibited consistently higher sensitivity in all scenarios.
Regarding study variables, demographic and comorbidity factors were employed, aligning with relevant literature. The challenge of handling missing values in blood test data was acknowledged, and future research could explore more suitable approaches for addressing this issue. It is recommended that future researchers undergo deeper training in database utilization, considering factors such as intraoperative metrics, to enhance the predictive capabilities of the model.
In conclusion, despite encountering challenges in database usage, the results of this study hold promise in providing practical insights for clinical practice, improving postoperative survival rates for elderly patients undergoing hip fracture surgery. Furthermore, by delving into the postoperative risks of elderly patients, we can better understand the medical needs of this population, offering valuable references for future healthcare services and policy formulation. |
