LASSO-Cox Modeling of Survival Using High-Resolution CT-Based Radiomic Features in a Cohort of COVID-19 Patients and Its Generalizability to Standard Image Reconstruction

Abstract

Background: Few studies have focused on predicting the overall survival (OS) of patients affected by SARS-CoV-2 (i.e., COVID-19) using radiomic features (RFs) extracted from computer tomography (CT) images. Reconstruction of CT scans might potentially affect the values of RFs. Methods: Out of 435 patients, 239 had the scans reconstructed with a single modality, and hence, were used for training/testing, and 196 were reconstructed with two modalities were used as validation to evaluate RFs robustness to reconstruction. During training, the dataset was split into train/test using a 70/30 proportion, randomizing the procedure 100 times to obtain 100 different models. In all cases, RFs were normalized using the z-score and then given as input into a Cox proportional-hazards model regularized with the Least Absolute Shrinkage and Selection Operator (LASSO-Cox), used for feature selection and developing a robust model. The RFs retained multiple times in the models were also included in a final LASSO-Cox for developing the predictive model. Thus, we conducted sensitivity analysis increasing the number of retained RFs with an occurrence cut-off from 11% to 60%. The Bayesian information criterion (BIC) was used to identify the cut-off to build the optimal model. Results: The best BIC value indicated 45% as the optimal occurrence cut-off, resulting in five RFs used for generating the final LASSO-Cox. All the Kaplan-Meier curves of training and validation datasets were statistically significant in identifying patients with good and poor prognoses, irrespective of CT reconstruction. Conclusions: The final LASSO-Cox model maintained its predictive ability for predicting the OS in COVID-19 patients irrespective of CT reconstruction algorithms.