以時間序列模型預測住院病人的生命徵象：模型間之比較分析

Abstract

本研究評估穿戴式智慧醫療裝置與時間序列預測模型於臨床生理監測中之實際應用與效能。基於目前普通病房面臨監測密度不足與人力缺乏之挑戰，我們希望透過智慧醫材結合時間序列模型強化早期預警能力，以減輕臨床醫護負荷。我們選用傳統統計模型、機器學習模型與近期發展之基礎模型，針對病人生命徵象數值進行回歸預測與預警趨勢分類任務之訓練與分析。本研究資料涵蓋台大醫院加護病房與普通病房，並包含四類異質性資料：加護病房生命監測器、加護病房穿戴式智慧醫療裝置、普通病房穿戴式智慧醫療裝置與台大醫院病歷系統紀錄。本研究對異質性資料採統一前處理流程，標準化時間格式、處理缺失與間斷資料，並建構滑動視窗架構，模擬連續監測情境下之預測任務。資料比對結果顯示，穿戴式裝置於心率指標表現出良好之趨勢一致性，在與加護病房監測器及病歷紀錄比對中皆具一定相關性，顯示其監測數據具備進一步進行時間序列分析之潛力。研究結果顯示，裝置種類與資料環境顯著影響模型表現，於回歸任務中，基礎模型在高頻資料中具備優勢，惟低頻或樣本數較小之資料，則顯示傳統模型表現亦不失穩定。另外，趨勢分類任務則因受限於臨床資料不平衡與指標設計，就上升趨勢辨識面臨重大挑戰。儘管如此，其初步結果已顯示智慧醫療裝置結合時間序列模型已具備臨床應用的潛力，為臨床於未來應用上，提供導入架構與即時模組的選項，實現高頻生命徵象監測之臨床應用。

This study evaluates the practical application and predictive performance of wearable devices integrated with time-series forecasting models for clinical vital sign monitoring. In response to the current challenges of insufficient monitoring density and limited clinical manpower in general wards, we aim to enhance early warning capabilities through the combination of smart medical devices and time-series models, reducing the workload of healthcare professionals. We utilized traditional statistical models, machine learning models, and recently developed time-series foundation models to perform both regression tasks and trend classification of patients’ vital sign trajectories. The dataset was collected from the intensive care unit (ICU) and general ward of National Taiwan University Hospital, comprising four heterogeneous data sources: ICU physiological monitor data, ICU wearable device data, general ward wearable device data, and structured vital sign records from electronic medical records. To handle the heterogeneity across datasets, we implemented a standardized preprocessing pipeline that included time format unification, handling of missing and discontinuous data, and a sliding window framework to simulate continuous monitoring conditions. Validation results showed that wearable devices, particularly in capturing heart rate, exhibited consistent trend alignment and a certain degree of correlation with both ICU monitors and EMR records, indicating their potential suitability for downstream time-series analysis. The results indicate that both device type and data environment significantly impact model performance. In regression tasks, foundation models exhibited superior performance on high-frequency data, whereas traditional statistical models also presented stable results in low-frequency or small-sample contexts. For the classification of NEWS (National Early Warning Score) trends, particularly for detecting elevated trends, challenges arose due to imbalanced clinical data and the discrete nature of the scoring system. Nonetheless, preliminary findings suggest that integrating wearable devices with time-series models holds clinical potential and may serve as the foundation for future implementations of real-time early warning systems for high-frequency vital sign monitoring.