The pandemic caused by SARS-CoV-2 continues to progress with over 23 million global cases and 800,000 deaths as of August 22, 2020. While the world awaits the emergence of an effective vaccine, a number of pharmacologic agents have been studied for COVID-19, the syndrome caused by SARS-CoV-2, but few have been shown to improve clinical outcomes.
The aim of this study is to describe outcomes, particularly mortality and clinical improvement, among patients with COVID-19 who receive different drugs (e.g. Tocilizumab and Remdesivir) within the Johns Hopkins Health System (JHHS).
We used different kinds of statistical tools including time-dependent propensity score matching, marginal structural model, and overlapping weights, to adjust for the bias of non-random treatment assignments. We compared the effectiveness of drugs and treatments to routine care among patients admitted to a 5-hospital health system. We also analyzed the side-effects of non-COVID treatments on patients with COVID.
Papers:
Elisa H. Ignatiusa# , Kunbo Wang# , et al, “Tocilizumab for the treatment of COVID-19 among hospitalized patients: A matched retrospective cohort analysis.”
link
Open Forum Infectious Diseases (2021).
Garibaldi Brian T.# , Kunbo Wang# , et al, “Comparison of Time to Clinical Improvement With vs Without Remdesivir Treatment in Hospitalized Patients With COVID-19.”
pdf
JAMA Open Network (2021).
Michael Burnim.# , Kunbo Wang# , et al, “The effectiveness of high-flow nasal cannula in COVID-19 pneumonia: a retrospective cohort study.”
link
Critical Care Medicine (2022).
Garibaldi Brian T.# , Kunbo Wang# , et al, “Real-world effectiveness of remdesivir in adults hospitalized with COVID-19: A retrospective, multicenter comparative effectiveness study.” Clinical Infectious Diseases (2021).
Garibaldi Brian T.# , Kunbo Wang# , et al, “Real-world effectiveness of remdesivir in adults hospitalized with COVID-19: A retrospective, multicenter comparative effectiveness study.” Clinical Infectious Diseases (2021).
Garneau William M., Liang Tao, Kunbo Wang, et al. “Clinical Outcomes of Patients Previously Treated with B-Cell Depletion Therapy Hospitalized with COVID-19: Results from the Johns Hopkins Crown Registry.”
Available at SSRN 4030554, (2022).
Boulanger Mary, Molina Emily, Wang Kunbo, et al. “Peripheral plasma cells associated with mortality benefit in severe COVID-19: A marker of disease resolution.”
The American journal of medicine, (2021).
In this project, we proposed a Reinforcement Learning framework to decide oxygen treatments on COVID patients dynamically at each decision point. Our model has advantages over the existing AI clinician models in two aspects. First, our proposed model has a built-in safeguard that involved clinical treatment guidelines to avoid unsafe treatment recommendations from RL model. Second, our model considered treatment stability at each decision time point. This is essential for oxygen treatments as it is unsafe in practice to change oxygen devices frequently.
Our clinician-involved RL model successfully avoided unsafe treatment recommendations, and performed better than clicians’ treatments on offline testing data.
In this project, we propose a Bayesian framework of regression model to predict a multidimensional array (tensor) of arbitrary dimensions from another multidimensional array of arbitrary dimensions. The framework is based on the contracted product of tensors, and Tucker decomposition of regression coeffcient tensor. Proper prior distributions are given to factor matrices of Tucker decomposition as well as core tensor, resulting in full posterior conditional distributions given in closed form formulas. Metropolis-Hastings method is used to choose the dimensions of core tensor. An alternative computing strategy is also given to speed up computation. We also compare our Bayesian Tucker-based tensor regression model with the newly developed CP regression model numerically on simulated data, real facial imaging data, and 3D motion data, and demonstrate the effectiveness of our model against CP model.
Paper in preparation:
In this project, we plan to first establish the second-order optimality for the tank-deficient Stochastic Block Model, which is a building block for establishing the second-order results for the more general Random Dot Product Graph model. We showed under proper generality conditions, when block probability matrix is rank-deficient, a Bayes estimator can be second-order optimal, in the sense that the corresponding asymptotic variance achieve the information lower bound given the knowledge of rank.
Paper in Preparation:
In this project, we propose a Bayesian framework that addresses several key biomedical questions in such settings. First, we want to understand how the longitudinal biomarker patterns vary based on subject baseline characteristics and inherent disease cure status in order to compute the personalized probability of an inherent cure. For the uncured subjects, we are interested in how the diagnosis time depends on baseline characteristics and whether we can identify a change point in the evolution of the biomarker before diagnosis.