Researchers from the University of Michigan Health System studied the use of machine learning to assess the prognostic utility of radiomic features for in-hospital COVID-19 mortality. The study was conducted over X-ray images of 3313 hospitalized patients, from which 3310 provided clear views of the lungs. Of those, 590 died and 20 were discharged to hospice. The findings of the study revealed several significant findings about predictability of demographic and clinical factors as well as radiomic texture features on in-hospital mortality of COVID-19 patients.
The predictive performance of five algorithms, including the Cox proportional hazards model, survival support vector machines, random survival forests, survival gradient boosting, and ensemble averaging of the first four algorithms was compared in the study. It was noted that ensemble averaging outperformed the other methods, with the highest average C-index of 81.0%. Subgroup analyses were conducted to examine which subgroups would benefit more with the added image features. The most important clinical features were found to be age, indications of fluid and electrolyte disorders, respiratory rate, diastolic blood pressure, metastatic cancer, and solid tumor cancer without metastasis. Important imaging texture features included dependence non-uniformity, zone entropy, median pixel intensity, large area high gray level emphasis, maximal correlation coefficient, pixel intensity kurtosis, and robust mean absolute deviation. Older age, higher respiratory rate and indications of fluid and electrolyte disorders, metastatic cancer, and solid tumor cancer without metastasis were all significantly associated with higher in-hospital mortality. Risk scores with and without radiomic features were constructed and it was noted that these scores could well distinguish patients across all the subgroups, reinforcing the usefulness of clinical and image features in profiling the risk of patient mortality.
Among the researchers, The University of Michigan Health System is a primary, regional center managing the care of patients with COVID-19 in the pandemic. It leverages machine learning techniques to extract, select and assess features from COVID-related X-ray images, playing an important role in providing valuable healthcare for patients. The person mentioned in this article is Dr. Jennifer Elisseeff, who is a professor of biostatistics at the University of Michigan School of Public Health and lead author in this study. She worked with Michigan Medicine’s Department of Radiology and Department of Biomedical Informatics to develop the study. This is an important and valuable research as it will help better understand and predict the mortality rate of COVID-19 patients. This research is also particularly useful for vulnerable patients, including the elderly and sick, who can better benefit with the imaging features.
