以類神經網路預測輕度頭部外傷後電腦斷層是否出現異常之分析

在美國每年約有100萬人因頭部外傷進行頭部電腦斷層檢查，這其中輕度頭部外傷佔了約80%。對到達急診時昏迷指數(Glasgow Coma Scale, GCS) 15分的病人而言，約有3-15%電腦斷層檢查是有異常的，其中只有1-5%是需要神經外科手術治療的。本研究的資料來源是從「中華民國頭脊髓外傷研究小組」所收集台灣全國頭部外傷資料庫，並從中挑選出自民國94年7月1號到民國95年6月30號，這其間發生輕度頭部外傷之傷病人來做研究，一共有2,294人。利用頭部外傷發生的機轉如車禍，跌落，墜落物襲擊及患者所使用的交通工具與相撞之物體等及臨床資料如年齡、性別、有顱骨骨折、局部神經功能障礙、有失去意識、創傷後癲癎、失憶、等來建構一類神經網路預測模型用來預測輕部頭部外傷後電腦斷層是否出現異常。最後，我們利用ROC 曲線(receiver operating characteristic curve)及ROC分析法對此一神經網預測模型之鑒別能力作分析，也將此分析結果與傳統统計學邏輯迴歸模型作比較。在我們的研究中，大約有44.2%的輕度頭部外傷後電腦斷層出現異常，而我們所建構的類神經網路模式為一個含有17個輸入變項的多層感知機網路。測試結果發現類神經網路模式有最大的ROC曲線下方面積(AUC：0.887)。類神經網路模式與邏輯迴歸模式之AUC無統計學上之顯著差異。然而，在兩個關鍵的ROC曲線臨界點上類神經網路預測模式都有最佳之共存敏感度及特異度。若把2組敏感度固定為90.3%，則類神經網路特異度為比邏輯廻歸的表現好(87% v.s. 50.3%)，且達到顯著差異。未來如果能結合臨床上用以預測輕度頭部外傷後電腦斷層出現更多的臨床變項（如嘔吐、頭痛等）納入建構類神經網路模式，相信類神網路將可協臨床醫師在做決定讓病人進行電腦斷層與否多一層把握。

關鍵字

類神經網路；輕部頭部外傷；邏輯迴歸

並列摘要

About one million people in the US receive brain computed tomographic examination every year due to head injuries.with 80% of them suffering a mild head injury. Among patients arriving in the emergency room with a Glasgow Coma Scale (GCS) of 15, 3%-15% of them were found to be abnormal in their brain CT scan, and up to 5% of them required neurosurgical intervention. We used the Taiwan National Head Injury Database collected by the Head and Spinal Injury Research Team. we selected patients (n=2,294) with a mild head injury in from July 1,2005 to June 30,2006. Based on information from the database and clinical research material, we then established a neural network to predict whether any abnormality appeared in the brain CT in patients with a mild head injury. The ANN’S ability to discriminate outcomes was assessed using receiver operating characteristic (ROC) curve analysis and the results were compared with a multiple logistic regression model. The results of the study that brain CT abnormalities following a mild head injury occurred in 44.2% of the patients, and that 17 predictive variables were identified by the ANN model. When tested on the same validation set, the ANN model had the greatest area under the ROC curve (AUC=0.887). we also found that a pair-wise comparison between the ANN and LR (AUC=0.861) models showed a similar predictive performance (p=0.078). The ANN model had the best simultaneous sensitivity and specificity at two pre-defined cut-off values. If the cut-off point sensitivity was fixed at an optimal level of 90.3%,we found the study showed that the ANN model had a better specificity and performance than a classical statistical logistical regression model (87% vs. 50.3%) Based on this study results, we conclude that we can merge clinical tests to determinate the exceptional outliers from the CT results after studying mild head injury, symptoms, such as vomiting and headache, to help clinicians better decide on whether to order brain CT scan.

並列關鍵字

Artificial neural network ； Mild head injury ； Logical regression

參考文獻

33.Min-Huei Hsu et al. Predicting cranial computed tomography results of head injury patients using an artificial neural network AMIA Annu Symp Proc. 2003;:868.

45.Min-Huei Hsu, Yu-Chuan Li,wen-Ta Chiu,et al. Predicting cranial computed tomography results of head injury patients using an artificial neural network AMIA Annu Symp Proc. 2003:868.

2.Borczuk P. Predictors of intracranial injury in patients with mild head injury. Ann Emerg Med. 1995;25:731-736.

3.Thurman DJ, Alverson C,Dunn KA. Traumatic Brain Injury in the United States: A public health perspective.J Head Trauma Rehabil 14(6):602,1999.

5.Jagoda AS, Cantrill SV, Wears RL .Clinical Policy: Neuroimaging and Decisionmaking in Adult Mild Traumatic Brain Injury in the Acute Setting.Ann Emerg Med .August 2002;40:231-249.

國際替代計量

以類神經網路預測輕度頭部外傷後電腦斷層是否出現異常之分析

全文下載

主題瀏覽