| 摘要: | 肺炎及肺水腫在急重症照護中，是臨床重症呼吸窘最常見的原因，呼吸道維持及保護心臟功能是這兩類病人搶救的基礎，因此在第一時間正確判斷是心臟衰竭或者肺部感染，並給予適當處置，不僅可能縮短病人住院的病程，更在極端年齡層如超過65歲的老人，以及多重共病問題的重症病人上，可以因為正確的決策方向挽救寶貴的生命；迄今，雖然許多實驗室檢查，可以輔助確立診斷，但是均不及X光片迅速方便(如：Portable X光片) ，然肺炎及肺水腫兩種疾病影像有許多極為類似相近之處，對第一時間要下處置之急重症醫師，是極困難的任務，且世界各地醫院放射專科醫師，皆無法全天在第一時間給予立即影像學的專門報告，因此如何在取得影像後，正確判讀是目前臨床嚴峻的挑戰。從過去的文獻當中已知深度卷積神經網絡(Deep convolutional neural networks：CNNs) ，對於胸部X光片有高度的辨識度，唯這類研究尚未在最急需照護的重症病人當中進行探討，這類的病人通常為極端年齡如：65歲以上之高齡病人，可能合併多種共病甚至臥床無法標準擺位，體內外又往往有多條儀器線路或者維生管路，這些干擾可能會影像機器學習。
本次研究回溯蒐集2015-2020五年，大於65歲以上因肺炎或肺水腫住院病人之前胸照X光影像，用ICD-10確定兩種疾病，並將同時包含J18(肺炎)及J81(肺水腫)之共病檔案刪除，再利用電子病歷中不同文字標籤分類與影像的相關性，將影像分做正相關、負相關、無相關、低相關和極相關並將不同特徵之影像放入不同子目錄，另外對將無干擾以及裁切干擾的影像進行分類，使用GoogLeNet Inception V4進行六次實驗以及三次驗證。
我們最好的實驗模型（實驗六），使用明確文字標籤後肺炎與肺水腫的 F1 score分別為0.835、0.829；模型準確率：83.2%、召回率：83.2、陽性預測率：83.3%以及F1 Score：0.832，不僅顯著優於無干擾以及經過裁切處理的影像分類；再與過去文獻模型比較後，實驗六模型在肺炎及肺水腫的各項數據均較為出色。
經過驗證我們注意到，雖然模型準確度最高可以到達73%，但仍不盡理想，不過我們發現模型對於不是肺水腫有很高的陰性預測率，可以判斷影像不是肺水腫；而在預測為肺炎時，有很高的陽性預測率來確定診斷；這部分本實驗模型可以作為臨床決策系統（Clinical Decision Support），用來作為排除肺水腫以及確診肺炎的工具，對老人急重症醫療做出貢獻。
Pneumonia and pulmonary edema are the most common causes of acute respiratory failure in emergency and intensive care. Airway maintenance and heart function preservation are two foundations for resuscitation. Not only can it reduce the hospitalization length, promptly diagnosing heart failure and pneumonia can also save patients' lives, especially for the elderly aged 65 and above and those with multiple preexisting conditions. Laboratory examinations have been utilized for clinicians to early differentiate pneumonia and pulmonary edema; however, none can provide results as prompt as radiology examinations, such as portable chest x-ray (CXR), which can quickly deliver results without mobilizing patients. However, similar features between pneumonia and pulmonary edema are found in CXR. It remains challenging for Emergency Department (ED) physicians to make immediate decisions as radiologists cannot be on-site all the time and provide support. Thus, Accurate interpretation of images remains challenging in the emergency setting. References have shown that deep convolutional neural networks (CNN) have a high sensitivity in CXR readings. To the best of our knowledge , there has no studies that examine CXR images in the critical setting with the interference of life-supporting catheters、instruments or with unsteady posture.
In this retrospective study, we collected the CXR images of patients over 65 hospitalized with pneumonia or pulmonary edema diagnosis between 2016 and 2020. After using the ICD-10 codes to select qualified patient records and removing the duplicated ones, we used keywords to label the image reports found in the electronic medical record (EMR) system. After that, we categorized their CXR images into five categories: positive correlation, negative correlation, no correlation, low correlation, and extreme correlation. Subcategorization was also performed to better differentiate characteristics. We applied six experiments includes the crop interference and non- interference categories by GoogLeNet Inception V4 and applied three times of validations.
In our best model (Experiment 6), the F1 scores for pneumonia and pulmonary edema are 0.835 and 0.829, respectively; accuracy rate: 83.2%, Recall rate: 83.2%, positive predictive value: 83.3%, and F1 Score: 0.832 after using extreme noise labeling category. The results show significantly positive than other experiments such as no interference model and cropping interference model. It is also better than the previous literature models such as CheXneXt.
After the validation, the best accuracy rate of our model can reach up to 73%. Although we were not satisfied with the accuracy rate, the model has a high negative predictive value of excluding pulmonary edema, meaning the CXR shows no sign of pulmonary edema. At the time, there was a high positive predictive value in pneumonia. In that way, we could use it as a clinical decision support (CDS) system to rule out pulmonary edema and rule in pneumonia contributing to the critical care of the elderly. |
