目的：加護病房相關研究中，臨床病患生命徵象預測一直是非常重要的議題。然而
過去電子健康記錄資料的分析，大多是以數值型資料為主，文字型資料的研究鮮少
被發表。
方法：根據 MIMIC-III 數據庫中所收集到的數值型資料及半結構化資料（病患的臨
床記錄、診斷資料及檢驗報告），利用自然語言處理技術中的隱含狄利克雷分佈
（LDA）模型與機器學習方法，對 38,597 名成年的加護病房患者進行臨床結果預
測。
結果：整合結構化資料及半結構化資料可增加預測之準確性，其中長期死亡率最佳
AUROC 達到 0.871，短期死亡率最佳 AUROC 更能達到 0.922。
結論：本研究構建之模型可明顯提升預測效能，並成功辨識出重要變數。期望這樣
的分析結果能增加醫護人員對於病患病情的掌握，也讓醫療資源能夠得到更優化的
應用。

Objectives: Predicting clinical patients’ vital signs remains a critical issue in intensive
care unit (ICU) related studies. However, studies on electronic health record (EHR) data
have mostly analyzed numerical data and rarely semi-structured textual data.
Methods: Our study used structured and semi-structured data (i.e., patients’ diagnosis
data and inspection reports) collected from the MIMIC-III database. First, we used
the Latent Dirichlet Allocation (LDA) model (a model employed in natural language
processing) to process semi-structured data. Then, we used machine learning methods for
the prediction of clinical outcomes in 38,597 adult ICU patients.
Results: Based on the results, combining the structured and semi-structured data of
ICU patients can strengthen the ICU patient mortality prediction accuracy. The model
with machine learning methods generated favorable mortality predictions, where the
highest AUROC, for long-term mortality is 0.871, and the highest AUROC for short-term
mortality is 0.922.
Conclusions: The constructed model successfully identified crucial variables for
predicting patient mortality. Thus, when providing medical services to patients, health care
personnel may consider the critical variables associated with the patients’ hospitalization
durations to ensure that the patients receive optimal medical services.