New Data Method Drives Hospital Care for COVID-19 Patients
Machine learning technology helps medical staff at Johns Hopkins hospitals predict which patients are most likely to experience severe COVID-19
I’ve built a career helping public officials think about how they can use vast data assets to make better decisions and improve conditions in their communities. Here at the PDI we have focused largely on the high level measures that are guiding policy makers through their decision-making process. But what does that feel like at the bedside, on the COVID units and in our ICUs? How are medical professionals using data to support their decision making as they also navigate the unknowns of a new virus? One example of this comes from Johns Hopkins inHealth, the precision medicine arm of Johns Hopkins, which seeks to leverage what has been learned from the COVID-19 pandemic to improve discovery and clinical outcomes.
inHealth’s Precision Medicine Analytics Platform (PMAP) and the JH-CROWN registry -- which includes anonymized data such as demographic characteristics, medical history, and laboratory results from COVID-19 patients at all five hospitals in the Hopkins system -- have enabled teams to develop new tools for patients and healthcare providers.
One data-driven invention to come out of this framework was the COVID-19 Inpatient Risk Calculator (CIRC) 1, which quantifies the probability of progression to severe disease or death among patients hospitalized with COVID-19. Researchers designed algorithms and statistical models to analyze and discover patterns in COVID-19 patient data. This machine learning process measured the historical effects of readily accessible patient data (demographics, vital signs, lab results, etc.) on their COVID-19 outcomes. The results were then assembled in a calculator that could compute the probability of a given patient progressing to severe disease once that individual’s data were inputted. Armed with support from the tool, medical teams could then help make critical decisions about where and how to treat patients.
The development of CIRC was driven by a desire to understand the risk factors that predict hospitalization and the likelihood of a patient progressing to severe disease or death. The team found that if a patient came from a nursing home, had other underlying conditions/illnesses, was older, had a higher BMI, etc., it was more likely that they would experience an undesirable outcome. However, CIRC’s power as a prediction tool was limited because it only used data from a patient’s admission to the hospital and could not update over time.
CIRC laid the groundwork for the subsequent development of the Severe COVID-19 Adaptive Risk Predictor (SCARP) 2, which does allow for real-time updates to predictions as new patient data are acquired. SCARP bins vital signs and labs into discrete chunks of time, accounts for all the bins and their changes, and then provides an updated prediction of COVID-19 prognosis. It can be used in real time by clinicians within the first 14 days of a hospital stay to identify patients with a very high likelihood of requiring imminent ICU-level care and plan interventions to alter that trajectory.
A version of SCARP has been available in the Johns Hopkins Electronic Health Record (EHR) since late January 2021. There is also a free version available online for providers to use the tool at hospitals outside of the Johns Hopkins network.
Because Johns Hopkins uses a universal data language, it can be transparent about the methods used to create models and to validate them with its proprietary data, then share the code. While privacy is always an issue with patient data, this system allows Johns Hopkins to make tools available for use at other hospitals without sharing specific patient data.
None of this data would exist without the sacrifices of frontline healthcare providers, patients, and their families. And, Hopkins researchers plan to continue working with these data to determine how the lessons learned from CIRC and SCARP might apply to diseases other than COVID-19. Researchers are adapting the risk evaluation methods established by CIRC and SCARP to similar respiratory diseases, such as influenza and pneumonia. The ultimate goal is for medical providers to walk into a patient’s room armed with data about everything that has happened to the patient up to that moment, gather additional information, and then use all available data to make the next decision. That is the promise of true precision medicine, and a bounty of skills that decision makers across the board should understand and begin to hone.
Title image courtesy of Dr. Shannon Wongvibulsin
References
B.T. Garibaldi, J. Fiksel, J. Muschelli, M.L. Robinson, M. Rouhizadeh, J. Perin, G. Schumock, P. Nagy, J.H. Gray, H. Malapati, M. Ghobadi-Krueger, T.M. Niessen, B.S. Kim, P.M. Hill, M.S. Ahmed, E.D. Dobkin, R. Blanding, J. Abele, B. Woods, K. Harkness, D.R. Thiemann, M.G. Bowring, A.B. Shah, M.-C. Wang, K. Bandeen-Roche, A. Rosen, S.L. Zeger, A. Gupta, Patient Trajectories Among Persons Hospitalized for COVID-19, Annals of Internal Medicine 174(1) (2020) 33-41.
S. Wongvibulsin, B.T. Garibaldi, A.A.R. Antar, J. Wen, M.-C. Wang, A. Gupta, R. Bollinger, Y. Xu, K. Wang, J.F. Betz, J. Muschelli, K. Bandeen-Roche, S.L. Zeger, M.L. Robinson, Development of Severe COVID-19 Adaptive Risk Predictor (SCARP), a Calculator to Predict Severe Disease or Death in Hospitalized Patients With COVID-19, Annals of Internal Medicine (2021).