| 摘要: | 背景 : 睡眠障礙涵蓋範圍廣泛，常見包含低氧血症(Hypoxemia)、阻塞性睡眠呼吸中止症（Obstructive Sleep Apnea；OSA）、猝睡症; 睡眠不足等，全球睡眠缺氧患者數量約10億人，佔全球總人口約12.6%，睡眠血氧濃度檢測是評估睡眠呼吸中止症的重要指標，但目前的檢測方法存在痛苦、昂貴、時間與地點受限等問題。
目的 : 本研究提出利用呼氣丙酮濃度分析替代血液或尿液酮體的檢測，以評估睡眠血氧濃度。呼氣丙酮是酮體的良好預測因子，利用呼氣丙酮濃度分析具有無創、頻繁、低成本等優點。
方法 : 本研究受試者條件，包含年齡介於20至60歲之間、性別不限、試驗前三個月無住院史，收案時間為期兩週，進而研究呼氣丙酮變化與睡眠血氧濃度關聯性，以驗證呼氣丙酮濃度分析是否可作為睡眠血氧濃度評估的有效工具。
結果 : 本研究結論為呼氣丙酮變化與睡眠血氧濃度高度關聯性，以本研究數據集進行機器學習建立二分類預測模型，經由五種模型(邏輯斯回歸、K-近鄰、決策樹、隨機森林、極限梯度提升)進行比較，得出隨機森林(Random Forest)模型預測效能最佳，準確率(Accuracy)達0.97±0.00、精確率(Precision)達0.87±0.01、召回率(Recall)達0.94±0.02、特異性(Specificity)達0.97±0.01、F1-Score達0.90±0.00。未來有機會藉由呼氣丙酮濃度分析方式，建立AI預測模型，提供醫療人員快速睡眠血氧濃度評估參考資訊。
Background: Sleep disorders encompass a wide range of conditions, including hypoxemia, obstructive sleep apnea (OSA), narcolepsy, and sleep deprivation. Globally, approximately 1 billion people suffer from sleep-related hypoxia, accounting for about 12.6% of the world's population. Monitoring blood oxygen saturation during sleep is a crucial indicator for assessing sleep apnea. However, current methods for such monitoring are often painful, expensive, and limited by time and location constraints.
Objective: This study proposes the use of breath acetone concentration analysis as an alternative to blood or urine ketone body detection for assessing sleep oxygen saturation. Breath acetone is a good predictor of ketone bodies, and its analysis offers non-invasive, frequent, and low-cost advantages.
Methods: The study's inclusion criteria for participants include individuals aged between 20 and 60, regardless of gender, with no history of hospitalization in the past three months. The recruitment period lasted two weeks. The study investigates the correlation between breath acetone changes and sleep oxygen saturation to verify whether breath acetone concentration analysis can serve as an effective tool for evaluating sleep oxygen levels.
Results: The study concluded that changes in breath acetone are highly correlated with sleep oxygen saturation. Using the study's dataset, a binary classification prediction model was built and compared across five models (Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest, and Extreme Gradient Boosting). The Random Forest model demonstrated the best predictive performance, with an accuracy of 0.97 ± 0.00, precision of 0.87 ± 0.01, recall of 0.94 ± 0.02, specificity of 0.97 ± 0.01, and F1-Score of 0.90 ± 0.00. In the future, breath acetone concentration analysis may establish AI predictive models, providing healthcare professionals with rapid assessment information for sleep oxygen levels. |
