- 期刊
應用機器學習建置成人加護病房急性譫妄早期預測模型
Machine Learning for Constructing a Model for Early Prediction of Acute Delirium in Adult Intensive Care Units
摘要
譫妄屬急性、突發意識混亂及認知功能障礙問題，若未能及時發現處置，將影響病人安全。目前國內尚無應用機器學習建置譫妄預測模型相關研究，因此本研究目的在利用大數據資料擷取特徵因子，經由機器學習建置急性譫妄預測模型，採病歷回溯性研究設計，納入2022年05月01日至12月31日入住成人加護病房病人，以每8小時譫妄評估作為一事件，擷取至預測時間點的累積數據以預測譫妄發生，共計54,595筆事件，提取12個特徵選擇，經由資料處理及近鄰演算法補值，利用極限梯度提升（eXtreme Gradient Boosting, XGBoost）、邏輯迴歸、隨機森林（Random Forest, RF）及決策樹四種機器學習模型進行分析，取80%資料為訓練集進行模型調教，20%為驗證集進行內部驗證，最終進行外部驗證、模型推論、效能評估、調整模型參數及再次訓練模型。結果發現訓練集以XGBoost為最高（AUC=0.85），內部驗證以RF準確率最高（AUC=0.86），外部驗證以RF及XGBoost準確率最高（AUC皆為0.86），但在敏感度和模型信心驗證下，XGBoost較好。本研究內部驗證及外部驗證結果，皆在0.85以上準確度，顯示以機器學習建置的譫妄預測模型，可運用於譫妄預測之輔助性決策，提供護理人員早期偵測到譫妄發生。
並列摘要
Delirium is an acute disorder of consciousness that, if not detected and treated promptly, can strongly impact patient safety. Currently, no research has been conducted in China on the use of machine learning to develop a delirium prediction model. This study employed extensive datasets to identify key predictive factors and construct an acute delirium prediction model by using machine learning. Methods This retrospective study included patients admitted to an adult intensive care unit from May 1 to December 31, 2022. Delirium was assessed every 8 hours, with cumulative data captured up to the prediction time point. A total of 54,595 assessment events were recorded, and 12 key features were extracted from data. After the data were processed and the nearest neighbor algorithm was used to complete the data, the dataset was divided into 80% for training and 20% for validation. Four machine learning models were employed: the XGBoost, random forest, logistic regression, and decision tree models. External validation, performance evaluation, parameter tuning, and retraining were conducted as necessary. Results: The random forest model had the highest accuracy in the internal validation, with an area under the receiver operating characteristic curve (AUC) of 0.85. In the external validation, the random forest and XGBoost models achieved the highest accuracy, each with an AUC of 0.86. In cross validation, the random forest model had the highest accuracy, with an AUC of 0.77. Conclusion: With accuracy exceeding 0.85 in both the internal and external validation, the machine-learning-based model for predicting delirium exhibits strong potential for aiding early detection and supporting clinical decision making in intensive care.