| 摘要: | 研究動機與目的
Oxaliplatin是第三代鉑金類(platinum)的化療藥，臨床上常用於轉移性腸胃道惡性腫瘤，然而癌症病人在接受oxaliplatin治療的療程中可能會出現嚴重過敏反應，導致療程終止甚至死亡的情形。過去曾有多數研究描述關於oxaliplatin發生嚴重過敏反應的病人特徵，但結果仍眾說紛紜。本方欲透過機器學習的方法來預測病人使用oxaliplatin後發生嚴重過敏反應的可能性，以提供個人化預防過敏反應的醫療建議，保障病人接受oxaliplatin治療期間的用藥安全。
重要研究方法
本研究將使用支持向量機(SVM)、隨機森林(RF)、深度類神經網路(DNN)，以及極限梯度提升(XGBoost)等四種機器學習模型，將數據中抽取70%來訓練資料集，另外30%作為測試資料集，透過多項病人特徵來預測最後病人使用oxaliplatin後發生嚴重過敏反應的可能性，並以Kappa值、敏感性、陽性預測率(PPV)、正確率、精確度、特異性、陰性預測率(NPV)、F1分數以及馬修斯相關係數(Matthews Correlation Coefficient, MCC)來確認最優模型。
研究結果
此三種機器學習的預測模型分析測試集結果除了敏感性之外，其餘PPV、正確率、精確度、特異性、NPV、F1分數以及MCC都以XGBoost模型最高，因此使用五倍交叉驗證，100次特徵訓練方法對XGBoost模型進行最佳優化。最後再針對第四~五級過敏反應篩選了13項重要特徵，並以XGBoost模型的gain特徵排序方法，進行500次混淆矩陣訓練，獲得AUC為0.862、PPV為0.930、正確率為0.974之模型。
結論
本研究發現僅使用13項病人的臨床特徵即可以XGBoost模型準確地預測施打oxaliplatin會發生嚴重過敏反應的病人，此研究結果作為化療注射前之評估輔助工具，並提供化療前置藥物施打與處方建議，讓病人能順利完成oxaliplatin化療療程，維持最佳治療與預後。
Purpose
Oxaliplatin is a third-generation platinum chemotherapeutic agent. It is commonly used clinically for metastatic gastrointestinal malignancies. However, patients may experience severe allergic reactions during the course of treatment with oxaliplatin, and which may lead to the termination of the course of treatment or even death. In the past, studies have described the characteristics of patients with severe allergic reactions to oxaliplatin, but the results are still controversial. This study aimed to predict the possibility of severe allergic reactions after the use of oxaliplatin through machine learning, so as to provide personalized medical advice for preventing allergic reactions, and to ensure the safety of patients during the treatment of oxaliplatin.
Methods
This study used three machine learning models such as support vector machine (SVM), random forest (RF), Deep neural network (DNN), and extreme gradient boosting (XGBoost). 70% of the data was used as the training data set, and the other 30% used as the testing data set. Multiple characteristics of patients who received oxaliplatin therapy were used to predict the possibility of a severe allergic reaction, and we used Kappa value, sensitivity, positive prediction rate (PPV), precision, accuracy, specificity, negative prediction rate (NPV), F1 Score and Matthews Correlation Coefficient (MCC) to confirm the optimal model.
Result
Except the sensitivity, the PPV, precision, accuracy, specificity, NPV, F1 score and MCC were all higher in the XGBoost models then RF and SVM models. Therefore, the 100 random sampling methods were used to build the most optimally models for XGBoost. Finally, 13 important features were screened for the severe allergic reactions, and 500 times of confusion matrix training were performed using the gain feature ranking method of the XGBoost model. Finally, a model with an AUC of 0.862, a PPV of 0.930, and a accuracy of 0.974 was obtained.
Conclusion
This study found that the XGBoost model can accurately predict patients with severe allergic reactions after the administration of oxaliplatin using only 13 clinical characteristics of patients. The results can provide a prescription suggests that may help patients to complete the course of oxaliplatin chemotherapy successfully and maintain the best status of their disease. |
