Performance and Robustness of Machine Learning-based Radiomic COVID-19 Severity Prediction

Manual assessment of CTs has shown great promise in determining COVID-19 (C19) severity, however it is laborious and subtle CT findings can be overlooked. To address these problems, we developed and analyzed the performance and robustness of a logistic-regression (LR) model in predicting C19 severity in a large public cohort of 1110 patients.

The dataset provider and radiologist used imaging and clinical data to classify patients as mild, moderate, or severe. For each CT, 107 radiomic features (RF) were extracted. Selected RF were combined into a LR model for distinguishing severe from mild and moderate cases. The models were trained and validated with AUC on both observers’ classifications. Sensitivity analysis of imaging parameters and cross-validation (CV) on the inter-observer classifications determined model robustness.

A single RF (gray-level co-cccurrence matrix-Correlation) was sufficient to predict mild from severe C19 with AUC_provider=0.85 and AUC_radiologist=0.74 (CV yielded AUCs≈0.80). In predicting moderate from severe C19, first-order-Median RF alone had sufficient predictive power of AUC_provider=0.65. The AUC_radiologist increased to 0.66 as the number of RF grew to 5 (CV yielded AUCs≈0.62). Study suggests that RF may be useful for identification of severe C19 cases.