Development and validation of predictive models for COVID-19 outcomes in a safety-net hospital population

Abstract

Objective To develop predictive models of COVID-19 outcomes, elucidate the influence of socioeconomic factors, and assess algorithmic racial fairness using a racially diverse patient population with high social needs. Materials and Methods Data included 7,102 patients with positive (RT-PCR) SARS-CoV-2 test at a safety-net system in Massachusetts. Linear and nonlinear classification methods were applied. A score based on a recurrent neural network and a transformer architecture was developed to capture the dynamic evolution of vital signs. Combined with patient characteristics, clinical variables, and hospital occupancy measures, this dynamic vital score was used to train predictive models. Results Hospitalizations can be predicted with an Area Under the ROC Curve (AUC) of 92% using symptoms, hospital occupancy, and patient characteristics, including social determinants of health. Parsimonious models to predict intensive care, mechanical ventilation, and mortality that used the most recent labs and vitals exhibited AUCs of 92.7%, 91.2%, and 94%, respectively. Early predictive models, using labs and vital signs closer to admission had AUCs of 81.1%, 84.9%, and 92%, respectively. Discussion The most accurate models exhibit racial bias, being more likely to falsely predict that Black patients will be hospitalized. Models that are only based on the dynamic vital score exhibited accuracies close to the best parsimonious models, although the latter also used laboratories. Conclusion This large study demonstrates that COVID-19 severity may accurately be predicted using a score that accounts for the dynamic evolution of vital signs. Further, race, social determinants of health, and hospital occupancy play an important role.

Social determinants of health and the prediction of missed breast imaging appointments

“…We employed non-linear and linear classifiers including random forest (RF) [ 28 ], XGBoost [ 29 ], and the regularized versions of support vector machine (SVM) [ 30 ] and logistic regression (LR) [ 31 ] using l 1 or l 2 -norm regularization.…”

“…RF and XGBoost are non-linear algorithms that are difficult to interpret (often involving hundreds of decision trees) but are useful because they may indicate what is the best classification performance one could obtain. We also employed custom linear classifiers, including the support vector machine (SVM) [ 30 ] and logistic regression (LR) [ 31 ] which can yield interpretable models. SVM constructs a hyperplane that separates the two classes to maximize the margin between samples while minimizing misclassification errors.…”

“…Here, the coefficients of the LR parsimonious model are of great importance. After standardizing the variables, a larger absolute coefficient suggests that the likelihood is more sensitive to this specific variable [ 31 ]. The sign of the coefficient indicates the direction of correlation with MIA.…”

“…It uses a linear regression model to approximate the logarithm of the odds of the dependent variable (outcome) [33]. The regularized versions of LR and SVM (using l 1 or l 2 -norm regularization) were considered to improve the robustness of these algorithms in the presence of noise and outliers [31]. We used open-source python packages (i.e., Scikit-learn [34] and Statsmodels [35]) to implement our predictive models.…”

Predicting use of Intensive Care Units during the COVID-19 Pandemic

“…Rees et al 16 consider the duration of stay of a COVID-19 patient in the ICU as crucial data for decision-making and establish that the average duration is from 4 to 21 days. Hao et al 17 develop predictive models for disease severity according to the in uence of socioeconomic factors and racial equity. In contrast to the main approaches in the literature, we predict the number of COVID-19 ICU beds at the regional level by considering data from PCR tests and measures taken to prevent spread in the region.…”