Risk assessment of progression to severe conditions for patients with COVID-19 pneumonia: a single-center retrospective study

Lijiao Zeng; Jialu Li; Mingfeng Liao; Rui Hua; Pilai Huang; Mingxia Zhang; Youlong Zhang; Qinlang Shi; Zhaohua Xia; Xinzhong Ning; Dandan Liu; Jiu Mo; Ziyuan Zhou; Zigang Li; Yu Fu; Yuhui Liao; Jing Yuan; Lifei Wang; Qing He; Lei Liu; Kun Qiao

The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.03.25.20043166 doi: medRxiv preprint COVID-19 pneumonia is a multistate disease with clinically relevant intermediate endpoint like severity onset. Most survival data analyses set the onset as the primary end point, and censor recovery or hospital discharge. However, when competing risks of severity onset are present, this analytical method induces bias. In this study, the risk of severe progression assessed without considering the competition would be overestimated because the patients who would never progress (those who discharged from hospital without progression) were treated as if they could progress. The extent of such bias and its adjustment by competing risks modeling have been evaluated in clinical virology and oncology research [10] [11] [12] [13] . We incorporated high-dimensional variable selection techniques into the competing risks modeling so that quantitative image features can be extensively evaluated according to their contribution to risk prediction. Our evaluation results showed that incorporating CT image can significantly improve the prediction performance as compared to those only based on demographical and clinical information (mean time-dependent AUC = 0.880 versus 0.824). In particular, such improvement was achieved with only one additional image feature, suggesting the importance of using multi-modality data in risk analysis.