A Deep Learning Radiomics Model to Identify Poor Outcome in COVID-19 Patients With Underlying Health Conditions: A Multicenter Study

Siwen Wang,Di Dong,Liang Li,Hailin Li,Yan Bai,Yahua Hu,Yuanyi Huang,Xiangrong Yu,Sibin Liu,Xiaoming Qiu,Ligong Lu,Meiyun Wang,Yunfei Zha,Jie Tian
DOI: https://doi.org/10.1109/jbhi.2021.3076086
IF: 7.7
2021-07-01
IEEE Journal of Biomedical and Health Informatics
Abstract:Objective: Coronavirus disease 2019 (COVID-19) has caused considerable morbidity and mortality, especially in patients with underlying health conditions. A precise prognostic tool to identify poor outcomes among such cases is desperately needed. Methods: Total 400 COVID-19 patients with underlying health conditions were retrospectively recruited from 4 centers, including 54 dead cases (labeled as poor outcomes) and 346 patients discharged or hospitalized for at least 7 days since initial CT scan. Patients were allocated to a training set (n = 271), a test set (n = 68), and an external test set (n = 61). We proposed an initial CT-derived hybrid model by combining a 3D-ResNet10 based deep learning model and a quantitative 3D radiomics model to predict the probability of COVID-19 patients reaching poor outcome. The model performance was assessed by area under the receiver operating characteristic curve (AUC), survival analysis, and subgroup analysis. Results: The hybrid model achieved AUCs of 0.876 (95% confidence interval: 0.752-0.999) and 0.864 (0.766-0.962) in test and external test sets, outperforming other models. The survival analysis verified the hybrid model as a significant risk factor for mortality (hazard ratio, 2.049 [1.462–2.871], P < 0.001) that could well stratify patients into high-risk and low-risk of reaching poor outcomes (P < 0.001). Conclusion: The hybrid model that combined deep learning and radiomics could accurately identify poor outcomes in COVID-19 patients with underlying health conditions from initial CT scans. The great risk stratification ability could help alert risk of death and allow for timely surveillance plans.
computer science, interdisciplinary applications,mathematical & computational biology,medical informatics, information systems
What problem does this paper attempt to address?
The problem that this paper attempts to solve is how to accurately identify high - risk patients who may reach adverse outcomes (such as death) through initial CT scans in COVID - 19 patients with underlying health conditions. Specifically, the researchers developed a hybrid model that combines the 3D - ResNet10 deep - learning model and the quantitative 3D radiomics model, aiming to predict whether COVID - 19 patients will have adverse outcomes. The research background indicates that COVID - 19 has led to a relatively high morbidity and mortality rate among patients with underlying health conditions. Therefore, it is crucial to develop an accurate prognostic tool to identify adverse outcomes in these cases. ### Overview of Research Methods 1. **Dataset**: A total of 400 COVID - 19 patients with underlying health conditions from 4 centers were included in the study, including 54 death cases (labeled as adverse outcomes) and 346 patients who were discharged or hospitalized for at least 7 days. The patients were divided into a training set (n = 271), a test set (n = 68), and an external test set (n = 61). 2. **Model Construction**: - **Deep - learning Model**: Based on the 3D - ResNet10 architecture, it is used to extract features from CT images and predict the probability of adverse outcomes. - **Radiomics Model**: Quantitative features are extracted from CT images. The optimal feature subset is selected through the recursive feature elimination (RFE) algorithm, and the probability of adverse outcomes is predicted using a multivariate logistic regression model. - **Hybrid Model**: The probability results of the deep - learning model and the radiomics model are combined through multivariate logistic regression to form the final prediction probability. 3. **Evaluation Metrics**: - The area under the receiver operating characteristic curve (ROC) (AUC), survival analysis, and subgroup analysis are used to evaluate the model performance. - The AUC of the hybrid model on the test set and the external test set are 0.876 (95% confidence interval: 0.752 - 0.999) and 0.864 (95% confidence interval: 0.766 - 0.962), respectively, which are significantly better than other models. - Survival analysis shows that the hybrid model is a significant risk factor, with a hazard ratio (HR) of 2.049 (95% confidence interval: 1.462 - 2.871, P < 0.001), and it can well divide patients into high - risk and low - risk groups. ### Conclusion The hybrid model developed in this study combines the advantages of deep - learning and radiomics, and can accurately identify high - risk patients who may reach adverse outcomes in COVID - 19 patients with underlying health conditions. The excellent risk - stratification ability of this model is helpful for timely warning of the risk of death and formulating appropriate monitoring plans, thereby optimizing the allocation of medical resources in a tense clinical environment.