| 摘要: | 中文摘要
論文名稱：慢波睡眠與心率變異性的AI模型之可行性研究
臺北醫學大學醫學院人工智慧醫療碩士在職專班
研究生姓名： 陳家祺
指導教授：蘇家玉 臺北醫學大學 教授
研究背景：傳統上評估慢波睡眠的情形需依賴睡眠多項生理檢查，這種檢查過程繁瑣對患者而言，頗為不便，這限制了一些患者接受睡眠檢查的意願。為解決此一問題，本研究目的在探討一種基於心率變異性的人工智慧模型，用以預測慢波睡眠。透過應用機器學習方法或深度學習技術，提供一種更為簡便且可靠的方法，來獲取和預測睡眠週期中慢波睡眠的實際分佈。
研究方法：本研究?回溯性研究。資料來源為衛生福利部的雙和醫院
睡眠中心，505位受試者的睡眠多項生理檢查記錄，以睡眠多項生理檢查記錄進行了連續的心率變異性、血氧飽和度和電腦自動偵測慢波，頻率為0.4Hz ~ 4Hz。然後根據睡眠技師判定的睡眠分期記錄，以是否出現慢波數據進行標記。模型的特徵包括平均心率、九個心率變異性指標、和二個血氧飽和度指標；然後將數據集分為80%的訓練集和20%的測試集，使用四種機器學習方法與三種基於深度學習的時間序列模型進行模型訓練。在模型評估過程中，採用5折交叉驗證來計算模型的平均準確率（Accuracy）、接收者操作特徵曲?下面積（AUROC）和精確率-召回率曲?下面積（AUPRC），用以評估在這三個指標上，表現最佳的模型。另外，通過田口方法有效地優化多組合模型的最佳超參數組合。
研究結果：第三組實驗為最佳。數據集共有271,125筆，包含非慢波睡
眠事件共計188,931筆及慢波睡眠事件共計82,194筆。經由5折交叉驗證後，取模型評估指標Accuracy，AUROC及 AUPRC分別計算平均值，共有8個組合模型達到預期目標。機器學習方法以隨機森林（Random Forest）模型，應用隨機過採樣（Random Oversampling）效果最佳，達到的評估指標，Accuracy = 86.78%、AUROC = 0.934、AUPRC = 0.953。深度學習方法以長短期記憶（LSTM）模型，應用SMOTE Oversampling為最佳，Accuracy = 84.07%、AUROC = 0.915、AUPRC = 0.933。
研究數據顯示，以心率變異性作為人工智慧模型的特徵，適合用於預測慢波睡眠。這不僅提升了居家檢查的可行性和患者的接受度，還增?了醫生進行長期監測的能力。
ABSTRACT
Title of Thesis：Feasibility Study of an AI Model for Slow-Wave
Sleep and Heart Rate Variability
Author：Chia-Chi Chen
Thesis advised by：Emily Chia-Yu Su
Taipei Medical University,
Professional Master Program in Artificial Intelligence in Medicine
Traditionally, assessing Slow-Wave Sleep relies on Polysomnography, a cumbersome and inconvenient procedure for patients, limiting some patients' willingness to undergo sleep studies. To address this issue, this research aims to explore an Artificial Intelligence model based on Heart Rate Variability indicators to predict Slow-Wave Sleep. By applying Machine Learning and Deep Learning techniques, a more convenient and reliable method is provided to obtain and predict the actual distribution of Slow-Wave Sleep during sleep cycles.
This retrospective study sourced data from the Sleep Center at Shuang Ho Hospital, Ministry of Health and Welfare, involving 505 participants with Polysomnography records. Continuous Heart Rate Variability, Oxygen Saturation , and automatic detection of Slow Waves, specifically Slow Waves (0.4Hz - 4Hz), were recorded. Sleep stages were marked based on whether Slow Waves data appeared, as determined by sleep technicians. The model's features included mean Heart Rate , nine Heart Rate Variability indicators, and two Oxygen Saturation indicators. The dataset was divided into an 80% training set and a 20% testing set, using four machine learning methods and three deep learning-based time series models for training. Model evaluation involved 5-fold cross-validation to calculate average Accuracy, Area Under the Receiver Operating Characteristic Curve (AUROC), and Area Under the Precision-Recall Curve (AUPRC), identifying the best-performing models. Furthermore, the Taguchi method was employed to optimize the best hyperparameter combination for multi-model ensembles.
Research Results: The third experiment was the best overall. The dataset comprises a total of 271,125 records, including 188,931 records of non-slow-wave sleep events and 82,194 records of slow-wave sleep events. Using 5-fold cross-validation, the average values of the model evaluation metrics Accuracy, AUROC, and AUPRC were calculated. There are eight ensemble models that have achieved the expected goal. Among machine learning methods, the Random Forest model with Random Oversampling performed the best, achieving the following evaluation metrics: Accuracy：86.78%, AUROC：0.934, and AUPRC：0.953. For deep learning methods, the LSTM model with SMOTE Oversampling was the best performer, achieving the following evaluation metrics: Accuracy：84.07%, AUROC：0.915, and AUPRC：0.933.
The research data indicates that using heart rate variability as a feature for artificial intelligence models is suitable for predicting slow-wave sleep. This not only enhances the feasibility of home examinations and patient acceptance but also improves the ability of doctors to conduct long-term monitoring. |
