應用人工智慧模型於加護病房床位需求預測之研究

Application of Artificial Intelligence Models to Predict ICU Bed Demand

在重症加護病房（ICU）中，輔助準確預測患者的臨床結果對於提升醫療決策效率和患者預後相當重要。然而，由於ICU數據的多樣性與不平衡性，如何選擇適合的人工智慧模型並提升其效能成為一項課題。本研究旨在比較多種人工智慧模型在不同ICU科別（外科ICU、神經科ICU、內科ICU）中的預測效能，並探討數據平衡策略對模型表現的影響，以期為臨床應用提供具體指引。本研究選用支持向量機（SVM）、邏輯迴歸（LogisticRegression）、K最近鄰演算法（KNN）、高斯單純貝氏分類器（Gaussian Naive Bayes）、多層感知器神經網路（MLP Neural Network）及極端梯度提升（XGBoost）等模型進行效能比較。結果顯示，XGBoost在三個ICU科別中表現最為穩定，曲線下面積（Area Under theCurve, AUC）均為0.82，是跨科別預測的最佳選擇；相較之下，Gaussian Naive Bayes表現最差，AUC僅為0.63-0.64，顯示其難以有效捕捉ICU數據特徵。MLP神經網路在神經科與內科ICU的效能較外科ICU顯著提升，AUC達0.76；Logistic Regression與KNN模型在神經科及內科ICU的表現（AUC=0.74）亦高於外科ICU，顯示不同模型對特定科別數據的適應性差異。
為進一步提升模型效能，本研究以XGBoost模型為主要分析工具，並探討數據平衡策略（原始數據、過採樣、欠採樣）對模型性能的影響。結果顯示，原始數據集普遍存在過擬合問題，而過採樣與欠採樣策略均能顯著改善模型的預測能力。過採樣有效平衡類別分佈，適合一般應用；欠採樣則在內科ICU與神經科ICU的數據中表現稍優，顯示其可能有助於增進不平衡數據預測能力。
本研究旨確認XGBoost在ICU數據跨科別預測中的穩定性與優越性，並強調數據平衡策略對模型效能的重要性，為人工智慧在ICU中的應用提供實踐依據。

In the intensive care unit (ICU), accurate prediction of clinical outcomes among patients is crucial for improving patient prognosis and the efficiency of medical decision-making. However, diverse and imbalanced ICU data create challenges in the selection and performance improvement of suitable artificial intelligence (AI) models. In this study, the predictive performance of various AI models was compared across surgical, neurological, and medical ICU department data. Additionally, the effects of data balancing strategies on model performance were evaluated to provide practical guidance for clinical applications.
The performance of support vector machines, logistic regression models, K-nearest algorithms (KNN), Gaussian Naive Bayes, multilayer perceptron neural networks (MLP), and Extreme Gradient Boosting (XGBoost) was evaluated. XGBoost was identified as the best choice for cross-department predictions, demonstrating the most stable performance across all ICU departments, with an area under the curve (AUC) of 0.82. Gaussian Naive Bayes performed the worst in capturing ICU data features, with an AUC ranging from 0.63 to 0.64. The MLP model performed better with neurological and medical ICU data than with surgical ICU data, with an AUC of 0.76. Logistic regression and KNN models also performed better with neurological and medical ICU data (AUC = 0.74) than with surgical ICU data, highlighting differences in adaptability to specific departmental data between the models.
To further improve model performance, the effects of data balancing strategies (original data, oversampling, and undersampling) on predictive accuracy were evaluated with the XGBoost model. Although general overfitting was observed in the original dataset, both oversampling and undersampling strategies considerably improved the model’s predictive ability. Oversampling effectively balanced class distribution and was suitable for general application. Undersampling performed slightly better with medical and neurological ICU data, demonstrating its potential to enhance predictions for imbalanced datasets.
The results of this study confirm the stability and superiority of XGBoost in crossdepartmental ICU data predictions, underscore the role of data balancing strategies in improving model performance, and provide practical evidence for AI application in ICU settings.