| 摘要: | 不當使用抗生素導致細菌抗藥性發展快速，而現行微生物檢驗方法提供細菌對抗生素有無抗藥性之抗生素藥敏檢驗（Antibiotic Susceptibility Test）需要數天，無法快速提供臨床醫師抗生素使用之指引，導致無法立即正確施用，為引發抗藥性危機之可能原因之一。近年來，臨床微生物檢驗已廣泛採用基質輔助雷射脫附游離飛行時間質譜儀（MALDI-TOF MS）進行菌種鑑定，不須繁雜且耗時的實驗，即能取得高準確度之菌種辨認。儘管已有研究表明可利用基質輔助雷射脫附游離飛行時間質譜儀進行細菌菌株之分型或辨識其抗藥性，然而，目前仍缺乏基於大量細菌質譜數據辨識抗藥性之預測系統。本研究主要目的為基於質譜數據與藥敏檢驗結果，利用機器學習方法建構辨識細菌抗藥性之預測系統，此系統首先整合合作醫院多年蒐集之細菌質譜與其抗生素藥敏檢驗報告數據，建置整合性資料庫，以利針對不同細菌建置模型，其中，為取得質譜之數值，進行質譜數據之前處理，接著，開發不同方法解決質譜峰值偏移問題，並且，藉由溶血性葡萄球菌之菌株分型（Strain Typing of Staphylococcus haemolyticus）與耐甲氧西林金黃色葡萄球菌（Methicillin-resistant Staphylococcus aureus）之辨識分析探究適用性，針對溶血性葡萄球菌菌株分型採用基於統計檢定的方法估計質譜峰值對齊時的參考光譜，並利用機器學習演算法建構菌株分型分類器，以隨機森林建構之模型達0.866之準確率；針對耐甲氧西林金黃色葡萄球菌辨識，對4858 株金黃色葡萄球菌，以分箱方法（Binning Method）解決質譜峰值之對齊，採用機器學習演算法建構辨識模型，獨立測試結果顯示受試者操作曲線下面積（Receiver Operating Characteristic Curve）達0.8450；最後，我們針對臨床常見六種細菌建構其抗藥性模型並建置預測系統，XBugHunter，此預測系統應用於臨床之結果顯示，辨識耐甲氧西林金?色葡萄球菌以及多重抗藥性鮑氏不動桿菌（Acinetobacter baumannii）準確率皆超過90%，相較採用現行抗生素藥敏試驗之流程，利用XBugHunter預測結果平均可減少35.72小時之處理時間，此外，在金黃色葡萄球菌菌血症患者中，採用XBugHunter之實驗組患者28天內死亡率之相對風險降低了38.4%，透過開發之辨識細菌抗藥性預測系統，有助於即時施用適當藥物，以達到減低死亡率、避免抗藥性，以及縮短住院天數等效益。;Improper use of antibiotics has led to the rapid development of antibiotic resistance. The current antibiotics susceptibility test (AST) which provides information of microorganism against antibiotics in clinical microbiology laboratory would spend several days. Unable to provide timely accurate prescription of antibiotics is a possible reason to the emergence of resistant bacteria. In recent years, matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) has been widely used in clinical microbiology laboratories for rapid bacterial species identification. High accuracy bacterial species identification can be achieved without complicated and time-consuming experimental steps. Although several studies have shown that MALDI-TOF MS can be used for rapid AST, it is lack of a predictive system for identifying bacterial resistance based on mass spectrometry data. The purpose of this study is to develop a predictive system using machine learning algorithms to identify antibiotic resistant bacteria based on MALDI-TOF MS data and AST reports. First, we integrated different databases including bacterial mass spectral databases, AST report databases, and strain types databases collected from our cooperative hospitals. To integrate and obtain the value of the mass spectrum, we preprocessed the mass spectrum data. Then, we developed three methods to deal with peak shifting problems, and we explored the applicability through analysis of strain typing of Staphylococcus haemolyticus and identification of methicillin-resistant Staphylococcus aureus (MRSA). In the strain typing of application, we employed a statistical test to estimate the reference spectrum when solving peak shifting problems. We then used different machine learning algorithms to construct strain typing classifiers. The accuracy of the classifier constructed by random forest algorithm was 0.866. In the identification of MRSA application, we used binning method to deal with peak alignment issues. A large scale dataset which is different from studies in the literature was then used to construct models for identification of MRSA. Similarly, we used several machine learning algorithms to train and test the models. From the experimental results, our best model based on random forest algorithm achieved maximum area under the receiver operating characteristic curve of 0.8450 on the independent testing dataset. Finally, from these experiences, we implemented a predictive clinical system, XBugHunter, to identify antibiotic resistant bacteria for six common bacterial infections. The accuracies of clinical deployment for identifying MRSA and multi-drug resistant Acinetobacter baumannii were all higher than 90%. Compared with the current process of AST, the use of XBugHunter prediction can reduce the processing time or turn-around-time by an average of 35.72 hours. In addition, among patients with Staphylococcus aureus bacteremia, the relative risk of mortality within 28 days of the experimental group using XBugHunter was reduced by 38.4%. The predictive system for identifying bacterial resistance can provide clinicians with instant predicted ASTs, which helps clinicians prescribe appropriate antibiotics to achieve the benefits of reducing mortality, avoiding antibiotic resistance, and shortening the days in hospital. |
