In JMIR medical informatics ; h5-index 23.0
BACKGROUND : Most of mortality of COVID-19 were from severe patients.
OBJECTIVE : Effective treatment of these severe cases remains a challenge due to a lack of early detection.
METHODS : A total set of 27 severe and 151 non-severe clinical and computerized tomography (CT) records from 46 COVID-19 patients (10 severe, 36 non-severe) was collected for building the model. Using a recently published convolutional neural network (CNN), we managed to extract features from CT images. A machine learning model which combines these features with clinical laboratory results was also trained.
RESULTS : Herein, we presented a prediction model, combining the radiological outcome with the clinical biochemical indexes, to identify the severe cases. The prediction model yields a cross-validated AUROC score of 0.93 and F1 score of 0.89, which improved 6% and 15%, respectively, from the models with laboratory tests features only. In addition, we developed a statistical model for forecasting severity based on patients' laboratory tests results before turning into severe cases, with an AUROC score of 0.81.
CONCLUSIONS : To our knowledge, this is the first report on predicting COVID-19 patient's severity progression, as well as severity forecasting, through a combination analysis of laboratory tests and CT images.
Zhu Fang, Li Daowei, Zhang Qiang, Tan Yue, Yue Yuanyi, Bai Yuhan, Li Jimeng, Li Jiahang, Feng Xinghuo, Chen Shiyu, Xu Youjun, Xiao Si-Yu, Sun Muyan, Li Xiaona