以LSTM模型預測抗藥性金黃色葡萄球菌之院內感染風險

Taipei Medical University Institutional Repository > 管理學院 > 大數據科技及管理研究所 > 博碩士論文 > Item 987654321/61403

請使用永久網址來引用或連結此文件: http://libir.tmu.edu.tw/handle/987654321/61403

題名:	以LSTM模型預測抗藥性金黃色葡萄球菌之院內感染風險 Use of LSTM model in Nosocomial Infection Prediction of Methicillin-Resistant Staphylococcus Aureus（MRSA）
作者:	謝一豪
貢獻者:	智慧數據應用產業碩士專班陳錦華莊秀文
關鍵詞:	抗藥性金黃色葡萄球菌;機器學習;長短期記憶模型;院內感染;羅吉斯回歸模型 Methicillin-resistant Staphylococcus aureus;MRSA;Machine Learning;Long short-term memory;Nosocomial infection;Logistic Regression
日期:	2021-07-06
上傳時間:	2022-03-03 11:36:53 (UTC+8)
摘要:	抗藥性金黃色葡萄球菌，簡稱MRSA（Methicillin-resistant Staphylococcus aureus），起因於抗生素的過度使用，使得其成為超級細菌，具有引起大量感染及適應各種環境的能力，是目前常見也棘手的院內感染之一。本研究以台灣北部某區域醫院作為研究對象，其中符合院內感染MRSA條件的確診病例66個，再從做過MRSA檢驗但無確診的病例中隨機篩選出330個，進行模型訓練與感染預測。整合國內外MRSA感染危險因子及專家建議，以住院用藥、ATC藥品分類、手術、醫療行為等危險因子進行分析，並以入院後前7天資料及前14天資料，進行機器學習與長短期記憶模型（LSTM）來預測疑似病人感染MRSA的機率，並比較不同變數之連續數值及類別化對於模型影響，機器學習部分表現最好的是使用入院後前14天資料與16個因子的決策樹，F1-score平均為0.34， Recall為0.47；LSTM則以入院後前14天資料與428個因子表現最佳，F1-score平均為0.62， Recall平均為0.78。將病人住院事件時間納入的LSTM模型，其表現優於無時間概念的決策樹模型。另透過羅吉斯回歸分析，發現針對胃腸道和新陳代謝、心血管系統、皮膚科、生殖泌尿系統和性激素、神經系統與呼吸系統的用藥，為影響MRSA感染的危險因子。臨床上取得細菌培養結果約需要2至5天，這期間足以形成MRSA院內感染，運用醫療數據與模型分析，可以在不增加醫療成本與人力的情況下，更及時的針對高風險的病人進行隔離與處理，進而達到降低院內傳播的可能性。 Methicillin-resistant Staphylococcus aureus（MRSA）it has become a drug-resistant super bacteria due to our overuse of antibiotics , consequently cause a large number of infections by adapting to various environments. The MRSA infection is one of the most common and difficult nosocomial infections. This research setting is a regional hospital in northern Taiwan. We found 66 cases infected with MRSA and randomly sampled 330 suspected cases that were screened for MRSA but were not confirmed. These data were used to train the model to predict the MRSA infection. Base on the integration of domestic and foreign MRSA infection risk factors and suggestions from experts, we selected risk factors such as inpatient medication, ATC drug classification, surgery, medical behavior, etc- , as the input features in machine learning and long short-term memory（LSTM）model to predict the probability of MRSA infection. We compared and discussed the effects between different variable types including continuous and categories of the model performance. The best performance of the machine learning model is the decision tree which used data from 14 days post-hospitalization with 16 factors, the average F1-score is 0.34, and the average Recall is 0.47; the best performance of the LSTM models is using data from 14 days post-hospitalization with 428 factors, the average F1-score is 0.62, and the average Recall is 0.78. The LSTM model embedding the event time of patient’s hospitalization has an outstanding performance than the decision tree model which is not including any variables with event time of patients. In addition, we found that medications for the Alimentary tract、metabolism、Cardiovascular system、Dermatological、Genito-urinary system and sex hormone、Nervous system and Respiratory system are important risk factors affecting MRSA infections , through the multiple logistic regression model analysis. It takes about 2 to 5 days to obtain the results of bacterial culture. This period is so long to cause MRSA nosocomial infection. Utilizing medical data and modeling technique to build a prediction model, it is possible to isolate and treat high-risk patients immediately without increasing medical costs and manpower, thereby reducing the possibility of in-hospital transmission.
描述:	碩士指導教授：陳錦華指導教授：莊秀文委員：許明暉委員：溫淑惠委員：李垣樟
資料類型:	thesis
顯示於類別:	[大數據科技及管理研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	651	檢視/開啟

檢視Licence

在TMUIR中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....