Validation of a deep learning, value-based care model to predict mortality and comorbidities from chest radiographs in COVID-19

Abstract

We validate a deep learning model predicting comorbidities from frontal chest radiographs (CXRs) in patients with coronavirus disease 2019 (COVID-19) and compare the model’s performance with hierarchical condition category (HCC) and mortality outcomes in COVID-19. The model was trained and tested on 14,121 ambulatory frontal CXRs from 2010 to 2019 at a single institution, modeling select comorbidities using the value-based Medicare Advantage HCC Risk Adjustment Model. Sex, age, HCC codes, and risk adjustment factor (RAF) score were used. The model was validated on frontal CXRs from 413 ambulatory patients with COVID-19 (internal cohort) and on initial frontal CXRs from 487 COVID-19 hospitalized patients (external cohort). The discriminatory ability of the model was assessed using receiver operating characteristic (ROC) curves compared to the HCC data from electronic health records, and predicted age and RAF score were compared using correlation coefficient and absolute mean error. The model predictions were used as covariables in logistic regression models to evaluate the prediction of mortality in the external cohort. Predicted comorbidities from frontal CXRs, including diabetes with chronic complications, obesity, congestive heart failure, arrhythmias, vascular disease, and chronic obstructive pulmonary disease, had a total area under ROC curve (AUC) of 0.85 (95% CI: 0.85–0.86). The ROC AUC of predicted mortality for the model was 0.84 (95% CI,0.79–0.88) for the combined cohorts. This model using only frontal CXRs predicted select comorbidities and RAF score in both internal ambulatory and external hospitalized COVID-19 cohorts and was discriminatory of mortality, supporting its potential use in clinical decision making.

Opportunistic detection of type 2 diabetes using deep learning from frontal chest radiographs

“…Negative controls, henceforth referred to as controls, were chosen to mimic the deployment environment, and because the DL model would be deployed on all adult CXRs, controls including all available CXRs in the date range above were utilized, along with cases of T2D. The labels for imaging training were based on ICD10 Hierarchical Condition Category (HCC) codes (2021 model 24) for six disease classes, including T2D, congestive heart failure, cardiac arrhythmias, morbid obesity, chronic obstructive pulmonary disease, and vascular disease 21 . Codes mapped to a category were binary encoded to 1 (True), and absent codes mapped to 0 (False), utilizing the most recent codes as of December 2021.…”

“…Research has already demonstrated how DL with abdominal computed tomography imaging can detect numerous biomarkers predictive of, for example, metabolic syndrome in asymptomatic adults 18 . Likewise, DL with chest radiography has been shown to predict future healthcare expenses, health disparities, and multiple comorbidities 19 – 21 . Because of limitations around BMI, we aimed to explore the use of a multitask DL model to detect prevalent T2D from ambulatory frontal CXRs in a large clinical dataset.…”

Lessons learned: Development of COVID-19 clinical staging models at a large urban research institution

“…Eventually, given the inability to extract DICOM images from JPEG, all Chest X-Rays (CXR) were extracted from the PACS system utilizing a scripted method (SikuliX, 2.0.2) and saved in an encrypted laptop as high-quality 24-bit JPEG files (1669 × 1538 to 3032 × 2520 pixels). CXR analysis used for one particular deep learning algorithm predicting comorbidities and clinical outcome details was established in published work [ 11 , 12 ]. A solution to this issue may be to use a web environment with optional de-identification of imaging data to facilitate data distribution within a hospital environment [ 33 ].…”

“…We successfully created models and compared our results with other published models. These models were published in journals [9][10][11][12] and presented at national conferences. However, translation of such models to the clinical setting proved difficult, partly due to the malalignment of incentives throughout our hospital organization.…”

“…In addition, we generated multiple predictive models, incorporated machine learning, leveraged natural language processing (NLP), and performed imaging analytics. Our work has been published in several peer-reviewed journals [9][10][11][12][13]. New goals and aims were developed throughout the project, including examining social determinants of health and comparing our findings with larger national and regional COVID-19 patient datasets (e.g., NC3 [14]).…”