以機器學習方法結合檢傷結構化資料與就醫主訴預測急診病患的住院需求：時間序列驗證研究

Abstract

急診壅塞與床位調度壓力使得處置去向（disposition）的早期判斷成為臨床與醫務管理的重要課題。若能在病人到院檢傷（triage）階段，利用可立即取得的資訊預測住院風險，將有助於早期風險分層、資源配置與流程優化。本研究旨在建立並評估以機器學習預測急診處置去向（返家 vs 住院）的模型，整合檢傷階段之結構化資料與就醫主訴（chief complaint）文字語意特徵，並以時間序列切割（temporal validation）檢驗模型於跨年度資料上的泛化能力。 本研究使用單一醫療體系之多院區的急診資料，抽樣納入 2016–2022 年完成檢傷且具明確處置標籤之就診紀錄共 n = 160,407 筆，資料依年份進行資料集切分(time-based split)，其中訓練集（2016–2020）n = 124,934、驗證集（2021）n = 17,055、測試集（2022）n = 18,418。以就醫之結果定義二元目標（0=返家，1=住院）。訓練集、驗證集與測試集之住院比例分別為20.0%、23.7%、 22.5%。特徵包含人口學資訊、生命徵象、就醫相關特徵、各項缺失指標、時間特徵、及院區/科別相關欄位等結構化變項；另將就醫時的主訴文字透過 bert-base-chinese 轉換為 768 維的 BERT 向量（取 [CLS] embedding），作為可額外添加之語意特徵。模型比較包含 Logistic Regression、Calibrated Linear SVM、Random Forest、Gradient Boosting、XGBoost 與 MLP Classifier。在評估指標方面，本研究以 AUROC 作為主要模型比較指標，並觀察不同決策閾值（threshold = 0.10–0.50）下報告之 Accuracy、Sensitivity、Specificity、PPV、NPV 與 F1-score，以了解不同臨床取捨下之模型行為；此外，為進行主要臨床決策情境的定量呈現，本研究最終採用 threshold = 0.25 作為二元分類之決策門檻，並以此門檻在驗證集與測試集上彙整混淆矩陣與衍生指標。 在僅使用結構化檢傷資料（structured data）時，AUROC 最高之模型為XGBoost，在測試集中 AUROC = 0.832（驗證集 AUROC = 0.831）；在納入主訴 BERT 向量後（structured data + chief complaint BERT embeddings），部分模型的預測力表現還能進一步提升，AUROC 最高之模型仍為XGBoost，其測試集 AUROC = 0.845（驗證集 AUROC = 0.849）。比起僅使用結構化檢傷資料，多納入主訴 BERT 向量對於不同模型的增益程度不一：線性模型（Logistic Regression 與Calibrated Linear SVM）提升效果最明顯；樹狀增強模型（XGBoost、Gradient Boosting）呈現小幅提升。相對地，Random Forest 與 MLP Classifier在納入主訴向量後測試集 AUROC 未見明顯提升或有時呈現下降。當採用 threshold = 0.25 評估時，最佳模型為XGBoost，其 Sensitivity、Specificity、PPV、NPV 分別為 0.705、0.818、0.529、0.905（AUROC = 0.845）。上述的結果顯示，主訴語意 BERT 嵌入向量所提供之高維稠密語意訊號在不同模型下具有不同可利用性，可能與模型歸納偏好（inductive bias）、訓練收斂與正則化設定，以及跨年度資料漂移（temporal shift / text drift）下的泛化難度相關。 總結而論，本研究基於真實世界的臨床醫療數據，建立了一個可重現且貼近臨床部署情境的急診處置去向預測框架，並以時間序列驗證方式，評估跨年度泛化能力。研究結果顯示，整體表現最佳的模型為 XGBoost；若聚焦於納入主訴語意特徵後的相對增益，則以線性模型進步最為明顯。未來研究可進一步探討跨院外部驗證、模型校準（calibration）與漂移監測（monitoring）、以及在臨床流程中的人機協作機制，以提升模型安全性、可解釋性與落地可行性。

Emergency department (ED) crowding and bed management pressure make early prediction of patient disposition an important issue in both clinical care and hospital operations. If the risk of hospitalization can be predicted at the triage stage using information immediately available upon ED arrival, it may facilitate early risk stratification, resource allocation, and workflow optimization. This study aimed to develop and evaluate machine learning models for predicting ED disposition (discharge home vs. hospitalization), by integrating structured triage data with semantic features derived from chief complaint text, and to assess model generalizability across calendar years through temporal validation. This study used ED data from multiple hospitals within a single healthcare system. A total of 160,407 ED visits with completed triage records and clearly defined disposition labels from 2016 to 2022 were included. The dataset was split chronologically into a training set (2016–2020, n = 124,934), a validation set (2021, n = 17,055), and a test set (2022, n = 18,418). The binary outcome was defined as disposition status (0 = discharge home, 1 = hospitalization). The hospitalization rates in the training, validation, and test sets were 20.0%, 23.7%, and 22.5%, respectively. Features included structured variables available at the triage stage, such as demographic information, vital signs, visit-related characteristics, missingness indicators, temporal features, and hospital- or department-related variables. In addition, chief complaint text was transformed into 768-dimensional BERT embeddings using bert-base-chinese with the [CLS] embedding as the sentence representation. The candidate models included Logistic Regression, Calibrated Linear SVM, Random Forest, Gradient Boosting, XGBoost, and MLP Classifier. Model performance was primarily evaluated using the area under the receiver operating characteristic curve (AUROC). To characterize model behavior under different clinical trade-offs, Accuracy, Sensitivity, Specificity, Positive Predictive Value (PPV), Negative Predictive Value (NPV), and F1-score were examined across decision thresholds from 0.10 to 0.50. For the main clinical decision scenario, a threshold of 0.25 was ultimately selected, and confusion matrices together with derived metrics were summarized on the validation and test sets. Using structured triage data alone, the best-performing model was XGBoost, with an AUROC of 0.832 on the test set (validation AUROC = 0.831). After incorporating chief complaint BERT embeddings, the predictive performance of some models improved further. XGBoost remained the best-performing model, achieving a test AUROC of 0.845 (validation AUROC = 0.849). Compared with structured triage data alone, the incremental benefit of adding chief complaint BERT embeddings varied across models. The improvement was most pronounced in linear models, including Logistic Regression and Calibrated Linear SVM, whereas tree-based boosting models such as XGBoost and Gradient Boosting showed only modest gains. In contrast, Random Forest and MLP Classifier showed little improvement or occasional declines in test-set AUROC after the addition of chief complaint embeddings. At the selected threshold of 0.25, the best model, XGBoost, achieved a Sensitivity of 0.705, Specificity of 0.818, PPV of 0.529, and NPV of 0.905 (AUROC = 0.845). These findings suggest that high-dimensional dense semantic signals derived from chief complaint BERT embeddings are utilized differently across model classes, possibly due to differences in inductive bias, optimization and regularization behavior, and the challenges of generalization under temporal shift and text drift. In conclusion, this study established a reproducible machine learning framework for ED disposition prediction using real-world clinical data under a deployment-oriented temporal validation setting. XGBoost achieved the best overall performance, whereas linear models demonstrated the most notable relative improvement after incorporating chief complaint semantic features. Future work should further investigate external validation across hospitals, model calibration and drift monitoring, as well as human–AI collaboration mechanisms within clinical workflows, in order to enhance model safety, interpretability, and real-world feasibility.