摘要: | 阻塞性型睡眠呼吸中止症(Obstructive Sleep Apnea, OSA)對患者及醫療系統造成了沉重的負擔。配戴陽壓呼吸器(Continuous Positive Airway Pressure, CPAP)為目前治療阻塞性型睡眠呼吸中止症的黃金治療方法,並有顯著療效,然而患者治療順從性低為其治療方法面臨之主要挑戰。一個有希望的解決方案是提前感知睡眠呼吸事件的發生,使陽壓呼吸器能提前相應的調整壓力,從而改善陽壓呼吸器的治療順從性,並讓患者願意持續接受治療。使用整夜睡眠檢查併陽壓呼吸器壓力檢定(CPAP Titration)之數據進行分析,可能較能反映出患者在家中治療的反應。本研究旨在建立一種機器學習方法,利用回溯性整夜睡眠檢查併陽壓呼吸器壓力檢定數據及心跳變異率(Heart Rate Variability, HRV)特徵分析,提前預測需調壓事件的發生。研究採用七種監督式機器學習方法包括邏輯回歸(Logistic Regression, LR)、K-近鄰演算法(K Nearest Neighbors, kNN)、支持向量機(Support Vector Machine, SVM)、隨機森林(Random Forest, RF)、梯度提升機(Gradient Boosting Machine, GBM)、長短期記憶(Long Short-Term Memory, LSTM)和InceptionTime,提前60秒預測需調壓事件的發生。數據前處理以60秒心電圖(Electrocardiography, ECG)為一片段計算出平均心跳變異率,並以1秒為單位往後推移,將心電圖訊號轉換為連續的心跳變異率數據。接著,數據以每60秒心跳變異率片段做特徵萃取,並以60秒為單位往後推移,計算出時間序列型心跳變異率特徵。模型效能評估結果顯示,時間序列型(time-sectional)機器學習模型InceptionTime表現最優異,準確率(Accuracy)達到75.07%,而ROC曲線下面積(Area Under Curve, AUC)達到75.70%。特徵重要性分析(permutation feature importance)分析結果,血氧濃度相關參數及頻域分析(frequency-domain)的心跳變異率特徵-nHF,分別排序為第一至第三影響模型效能表現之最重要特徵。本研究結果顯示,在陽壓呼吸器壓力檢定過程中,透過血氧濃度參數及心跳變異率特徵,可以提前預測需調整壓力事件的發生。為未來改善陽壓呼吸器感測需調壓時機,以及提升居家使用陽壓呼吸器治療的耐受性及順從性,提供了一種新穎和有前景的方法。 Obstructive Sleep Apnea (OSA) imposes a significant burden on patients and healthcare systems. Continuous Positive Airway Pressure (CPAP) therapy is the gold standard treatment for OSA, providing substantial therapeutic effects. However, low treatment adherence poses a major challenge for patients undergoing CPAP therapy. An encouraging solution is the early prediction of sleep breathing events, enabling CPAP devices to proactively adjust pressure and improve treatment adherence, thereby promoting patient acceptance of the therapy. Analyzing data from overnight CPAP titration, may better reflect patients' responses to home-based therapy. This research aims to develop a machine learning approach utilizing retrospective data from overnight sleep studies – Polysomnogram (PSG), CPAP titration, and heart rate variability (HRV) analysis to predict impending pressure adjustment events. Seven supervised machine learning methods, including Logistic Regression (LR), K-Nearest Neighbors (kNN), Support Vector Machine (SVM), Random Forest (RF), Gradient Boosting Machine (GBM), Long Short-Term Memory (LSTM), and InceptionTime, were employed to predict pressure adjustment events 60 seconds in advance. Data preprocessing involved segmenting electrocardiogram (ECG) data into 60-second a 60-second window with a 1-second stride, and this rolling calculation aimed to determine continuous variations in HRV parameters. Following that, time-series data were produced using a 60-second window with a 60-second stride. Thereafter, SpO2-std (SD of the SpO2) and 7 continuous HRV features ((SDNN, RMSSD, NN50, HR-mean, nLF, nHF, nVLF)) for each 60-second window were determined. The performance evaluation of the models indicated that the time-sectional machine learning model, InceptionTime, outperformed others, achieving an accuracy of 75.07% and an Area Under the Curve (AUC) of 75.70%. Feature importance analysis revealed that blood oxygen concentration-related parameters and frequency-domain HRV features (e.g., normalized high-frequency power - nHF) were the top three most critical features influencing model performance. The results suggested that during CPAP titration, early prediction of pressure adjustment events can be achieved by leveraging blood oxygen concentration parameters and HRV features. This novel and promising approach provides insights for improving the detecting of pressure adjustment needs in CPAP devices and enhancing the tolerability and adherence to home-based CPAP therapy. |