Integrating plasma cell‐free DNA with clinical laboratory results enhances the prediction of critically ill patients with COVID‐19 at hospital admission

• Published in: Clinical and translational medicine 2022-07, Vol.12 (7), p.e966-n/a
• Main Authors: Bai, Yong, Zheng, Fang, Zhang, Tongda, Luo, Qiuhong, Luo, Yuxue, Zhou, Ruilong, Jin, Yan, Shan, Ying, Cheng, Jiehui, Yang, Zhimin, Li, Lingguo, Zhang, Haiqiang, Zhang, Yan, Yin, Jianhua, Fang, Mingyan, Chen, Dongsheng, Cheng, Fanjun, Jin, Xin
• Format: Article
• Language: English
• Subjects: Coronaviruses; COVID-19; Disease; DNA methylation; Genomes; Illnesses; Injuries; Laboratories; Letter to Editor; Medical prognosis; Plasma; Severe acute respiratory syndrome coronavirus 2; Survival analysis
• ISSN: 2001-1326; 2001-1326
• DOI: 10.1002/ctm2.966

Dear Editor, Owing to the substantial clinical heterogeneity of patients infected with SARS-CoV-2,1,2 factors primarily relying upon clinical and/or laboratory parameters are yet inadequate to accurately predict COVID-19 patients evolving to severe or critical illness at early stage.3,4 Recent studies have revealed an elevated level of cell-free DNA (cfDNA) in plasma in severe COVID-19 patients due to massive cell death or irreversible multiorgan injuries during pathological conditions.5,6 Therefore, the utilization of cfDNA profiling may benefit improving the COVID-19 prediction and help understand molecular characteristics of the life-threatening disease.7,8 Herein, we developed an M2Model, a LightGBM-based9 machine learning model with focal loss as an objective function to predict critical COVID-19 at admission by jointly analysing multimodal data, including laboratory parameters and cfDNA profiles. [...]the M2Model outperformed other single-type feature-based models in discriminating critical from noncritical COVID-19, achieving the highest AUROC (area under ROC curve) of .955 ± .029 (mean ± SD; Figure 1D) and AUPR (area under precision-recall curve) of .827 ± .153 (p < .0001; Figure 1E). Decision curve analysis and confusion matrix also demonstrated the superior prediction ability of the M2Model over other models (Figures 1G,H and S3A–D), with sensitivity of 85.19% (95% confidence interval [CI], 63.6%–100.0%), specificity of 93.33% (95% CI, 86.2%–98.6%), PPV (positive predictive value) of 66.67% (95% CI, 48.8%–88.9%), NPV (negative predictive value) of 97.58% (95% CI, 94.0%–100.0%) and MCC (Matthews correlation coefficient) of 71.02% (95% CI, 49.8%–88.8%) (Table S4).