機器學習用於血液透析中低血壓的檢測與分析

Abstract

本研究旨在使用傳統機器學習（Machine Learning）和深度學習（Deep Learning）模型開發一個透析中低血壓（Intradialytic Hypotension）的預測系統。數據集收集自台大醫院透析中心，涵蓋從 2016 年 1 月 1 日至 2023 年 9 月 30 日的透析記錄、生命徵象數據、患者資訊和醫療處置單。我們採用了序列特徵選擇（Sequential Feature Selection）和兩個門檻選擇標準來優化模型性能。在傳統ML模型中，XGBoost 和隨機森林均表現強勁，而在 DL 模型中，多層感知器（Multilayer Perceptron）則展現了優越的性能。我們的研究結果表明，這兩種類型的模型在預測 IDH 方面具有很強的能力。未來的工作將專注於動態調整窗口大小、整合歷史透析記錄以及開發針對個別患者定制的預測模型，以增強在實際臨床中的應用性並改善患者的治療效果。

This study aims to develop a predictive system for intradialytic hypotension (IDH) using conventional machine learning (ML) and deep learning (DL) models. The dataset, collected from the National Taiwan University Hospital dialysis center, includes dialysis records, vital sign data, patient information, and medical order records from January 1, 2016, to September 30, 2023. We employed Sequential Feature Selection (SFS) and two criteria for threshold selection to optimize model performance. Among conventional ML models, both Extreme Gradient Boosting (XGBoost) and Random Forest demonstrated strong performance, while Multilayer Perceptron (MLP) demonstrated superior performance among DL models. Our findings indicate strong predictive capabilities for both model types. Future work will focus on dynamic window size adjustments, integrating historical dialysis records, and developing customized prediction models to enhance real-world clinical applicability and improve patient outcomes.