重新定義庫欣氏反應以推估神經重症病患預後

Abstract

神經外科及重症醫師在處理神經重症病患時，常以顱內壓監測作為瞭解病人病情、調整藥物劑量及預後判斷之依據；但根據臨床經驗，顱內壓高低未必能夠準確反應病人真實情況。本研究利用大數據分析技術，包括小波轉換、線性迴歸與決策樹等，將顱內壓與其他生理參數如呼吸、血壓、心跳等結合，訓練出一組預測模型，以用於其他病人之評估。我們提出一個嶄新的概念，將傳統的庫欣氏反應 (Cushing response) 區分成病理性、生理性與陰性三類反應，擴大顱內壓對各種生理現象的調控機制，並藉由取得床邊監視器提供的生理訊號，經小波轉換以去除多數雜訊、線性迴歸以壓縮資料後，輸入決策樹模型，得到病程當中各時間點的狀態預測，與真實情況相比較，獲得很好的相關性，並藉由預測出來的庫欣氏反應分群比例，構成病況判斷與病程預測之警示系統。本模型預測準確率可達81.6%。未來目標則希望以非侵入性之生理參數來預測顱內變化，提供臨床醫師做為診斷、處置及病情解釋之參考；若能取代侵入性的顱內壓監測更佳。

In their daily care of neurocritical patients, neurosurgeons and intensivists often use intracranial pressure (ICP) monitoring to understand the neurological condition of the patient, to adjust their medications, and to evaluate the prognosis. In our experience, the absolute values of ICP cannot represent the clinical situation precisely. Our study uses big data techniques such as wavelet transform, linear regression and decision tree to combine ICP and other physiological parameters such as respiratory rate, heart rate, and arterial blood pressure and establish a predictive model. We propose a novel concept to divide Cushing response into pathological, physiological and negative ones to elaborate the control of pathophysiological mechanisms by ICP. All parameters are preprocessed with the wavelet transform, thus eliminating most noises, and compressed by linear regression. After random sampling from the transformed data, a decision tree is trained to produce a predictive model whose results are well compatible with clinical situations. The model is cross-validated and the accuracy of prediction evaluated by a confusion matrix. Based on the decision tree model, a warning light system is created to predict the situation and outcome of test patients. The accuracy of outcome prediction through the model reached 81.6%. In the future, we hope to establish the model with noninvasive physiological parameters to provide advice for clinical management and possibly to replace invasive ICP monitors.