| 摘要: | 背景：慢性硬腦膜下血腫疾病於65歲以上高齡患者有50%-80%病患有頭部外傷的病史，標準治療方式是進行血腫清除手術，手術成功率高但有0%-32%死亡率。過去文獻多以回歸分析方法提出此疾病的相關風險因子，但少見有應用機器學習做此疾病術後的存活預測模組。
目的：
(1) 、探討術後死亡原因
(2) 、建立存活預測模型
(3) 、比較不同模型的預測結果
(4) 、了解住院診斷關聯群（Taiwan’s Diagnosis-Related Groups, TW-DRGs）是否具備存活預測價值
方法：使用北醫三院資料庫醫療資料共766位成人，80%資料作訓練集，20%資料作測試集，使用SAS軟體進行統計資料分析、使用Python進行特徵選擇、建立機器學習預測模型（預測院內存活、預測六個月存活、預測一年存活）、並以SHAP value解釋各特徵對預測結果的貢獻。
結果：研究發現患有合併症和共病症的死亡病患其死因以惡性腫瘤、心臟疾病、腦血管疾病居前三名，慢性病占全部死因87%；另外，術前住院時間長，營養狀況不佳與高血糖病患也是術後死亡率高的危險因子。在模型預測結果中，預測院內存活結果最好的是Logistic Regression模型，其AUC 0.8893、Accuracy 0.69、Sensitivity 1、Specificity 0.78；預測六個月存活結果最好的是SVM演算法，其AUC 0.8197、Accuracy 0.86、Sensitivity 0.79、Specificity 0.83；預測一年存活結果最好的是KNeighborsClassifier演算法，其AUC 0.7962、Accuracy 0.8、Sensitivity 0.85、Specificity 0.75；在所有特徵中，三種特徵篩選方式於三種預測模型中均篩選出的特徵依序為：BUN、住院診斷關聯群、Glucose、PT、住院日數，顯示住院診斷關聯群（Taiwan’s Diagnosis-Related Groups, TW-DRGs） 除原本單純的批價申報碼以外，在python機器學習演算法中，對模型預測具有其相當重要性。
Background: Patients who are elder than 65 years old and suffering Chronic subdural hematoma disease and having head trauma are 50% to 80% of all. The standard treatment is to perform hematoma removal surgery. The success rate of surgery is high but there is a mortality rate of 0% to 32%. In the past, the literature has mostly proposed the relevant risk factors of this disease by regression analysis, but it is rare to have a postoperative survival prediction module for this disease using machine learning.
Objective:
(1) To investigate the causes of postoperative death
(2) To establish a survival prediction model
(3) To compare the prediction results of different models
(4) To understand whether Taiwan’s Diagnosis-Related Groups, (TW-DRGs) have considerable importance for model prediction
Methods: A total of 766 adults were used in Taipei Medical University Clinical Research Database (TMUCRD). 80% of the data was used as the training set, 20% of the data was used as the test set, and the statistical data was analyzed by SAS software. The main method is using Python for feature selection, building machine learning prediction models (predict in-house survival, predict six-month survival, predict one-year survival), and interpreting the prediction results of each feature sets with SHAP value.
Results: It was found that the causes of death of patients with comorbidities and common diseases were malignant tumors, heart diseases and cerebrovascular diseases, and all chronic diseases accounted for 87% of all deaths. In addition, patients with long preoperative hospital stays, poor nutritional status and hyperglycemia are also risk factors for high postoperative mortality. Among the model prediction results, the best prediction model Ⅰ (predict in-hospital survival) is the Logistic Regression model, which has AUC 0.8893, Accuracy 0.69, Sensitivity 1, and Specificity 0.78. The best prediction model Ⅱ (predict six-month survival) is the SVM algorithm, with AUC 0.8197, Accuracy 0 86, Sensitivity 0.79, and Specificity 0.83. The best prediction model Ⅲ (predict one-year survival) is the KNeighborsClassifier algorithm, which has AUC 0.7962, Accuracy 0.8, Sensitivity 0.85, and Specificity 0.75. Among all the features in the three feature selection methods of the three predictive models were screened out in order: BUN, Taiwan’s Diagnosis-Related Groups (TW-DRGs), Glucose, PT, and number of hospital days, showing the Taiwan's Diagnosis-Related Groups, TW-DRGs has considerable importance for model prediction in python machine learning algorithms. |
