Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/7296
Title: | 以支持向量機預測睡眠呼吸中止症 Support Vector Machine Prediction of Obstructive Sleep Apnea |
Authors: | Wen-Chi Huang 黃文圻 |
Advisor: | 賴飛羆(Feipei Lai) |
Keyword: | 呼吸中止症,支持向量機,機器學習,特徵篩選,預測系統, Obstructive Sleep Apnea,Support Vector Machine,Machine Learning,Feature Selection,Prediction System, |
Publication Year : | 2020 |
Degree: | 博士 |
Abstract: | 睡眠多項生理檢查 (PSG) 是現今阻塞型睡眠呼吸中止症 (OSA) 的標準診斷工具,然而進行PSG時,受測者需要配戴多種感測器,於睡眠檢查室進行一整夜的檢查,既耗時又不舒適。除此之外,非睡眠專科醫師在門診時,多會將所有疑似有睡眠障礙的病人,轉診至睡眠科,或安排PSG進行診斷,皆可能造成OSA診斷效率下降。為解決上述問題,已有多項相關研究使用非連續量測資訊來設計OSA的快速篩檢工具,以達到更有效率的PSG排程。然而在以往的研究中,多有靈敏度 (sensitivity) 高,特異度 (specificity) 低的問題,且在資料收集時,部份資料定義不夠明確,造成使用上的困難。本研究使用個人基本資料、身體量測資訊、共病症及睡眠症狀等32項候選特徵,為非睡眠專科醫師設計一項OSA的快速篩檢工具。本研究比較了八種方法的分類效能,並選擇其中結果最佳的支持向量機 (SVM) 來進行特徵篩選與最佳化。本研究提出了二個階段的特徵篩選,再使用SVM各自對三種不同嚴重程度的OSA [呼吸紊亂指數 (AHI) ≥5/hr、AHI ≥15/hr、AHI ≥30/hr] 進行二元分類器的訓練。特徵篩選的過程中發現,為達到較佳的分類效果 (AUROC ≥0.80),當預測三種嚴重程度的OSA時,分別需要2、6及6個特徵。在6,875人的資料中,使用2個特徵進行預測、針對 AHI ≥5/hr 的OSA患者,分類器效能(精準度、靈敏度、特異度)達到 (75.7%、76.4%、72.2% ),使用6個特徵進行 AHI ≥15/hr 的預測效能為 (73.7%、75.1%、70.4%),使用 6個特徵進行AHI ≥30/hr 的預測效能為 (73.0%、74.9%、70.0% )。本研究最終採用6個特徵做為快篩系統的輸入,並將訓練好的三個分類器,結合RedCap來提供網頁化的問卷,可即時回饋問卷填寫者OSA預測的結果,並提供醫師資料收集及閱覽的功能。 Polysomnography (PSG) is the gold standard for diagnosis of obstructive sleep apnea (OSA), but it is costly and access is often limited. Besides, the non-sleep specialist physician (NSSP) usually transfers most of suspected patients with sleep disorder to department of sleep for PSG. This situation lets OSA diagnosis more inefficient. To solve above problems, there were several studies used discretely objective and subjective information to develop OSA screening tools which provide diagnosis support and prioritize PSG. However, several OSA prediction models from recent studies had two issues: one is sensitivity range was higher than specificity range, and another is feature definitions as model input may be not clear enough. The first issue may lead more non-OSA patients to do PSG test, and the second may cause the difficulty of applying the prediction model. This study proposed an OSA screening tool for NSSP by 32 features which includes patient basic information, anthropometrics, comorbidities, and sleep habitual information. After comparing the performance of 8 algorithms, the support vector machine (SVM) had the best performance and was chose to be optimized with feature selection. The proposed method of this study applied a two stages of feature selection and using support vector machine to train classifier for OSA prediction. There were three classifiers trained for three apnea-hypopnea-index (AHI) cutoff (5/hr, 15/hr, and 30/hr). This study discovered that the classifier required more features with larger AHI cutoff to reach AUROC ≥0.80. Three AHI cutoffs required 2, 6, and 6 features, respectively. Three classifiers were trained and tested with 6,875 subject data. With 2, 6, and 6 features as input to predict three AHI cutoffs (5/hr, 15/hr, and 30/hr), the performance (accuracy, sensitivity, and specificity) achieved (74.24%, 74.14%, and 74.71%), (72.64%, 75.18%, and 68.73%), and (70.28%, 70.26%, and 70.30%), respectively. Finally, 6 features were selected and used in a web-based system which integrated trained SVM models. The web-based system is capable of giving user OSA risk as feedback in real-time, and it also lets medical staff collect and review the input data and outcome results. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/7296 |
DOI: | 10.6342/NTU202000475 |
Fulltext Rights: | 同意授權(全球公開) |
Appears in Collections: | 資訊工程學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-109-1.pdf | 2.32 MB | Adobe PDF | View/Open |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.