The Research of Predicting Septic Shock

How computational medicine is changing critical care in 5 questions

Participating in Research

As a new school year approaches, many medical students are opting to take a gap year dedicated to research. This trend is unique for students not in MD/PhD programs in the USA who have a deep interest in understanding and participating in research. A popular emerging field for the future of health care and medicine, known as computational medicine, is become an integral part of patient care. Regardless of location, students, as well as interns and health care professionals around the globe who are interested in emergency and critical care medicine, should consider this unique area of study as a part of their research gap year.

In this blog entry for the International Emergency Medicine Education Project (iEM), I discuss the role of computational medicine in detecting sepsis, one of the most important diagnoses to detect early, with Professor Rai Winslow, Director of the Institute for Computational Medicine at The Johns Hopkins University. As outlined on the Institute’s website, computational medicine “aims to improve health care by developing computational models of disease, personalizing these models using data from patients, and applying these models to improve the diagnosis and treatment of disease.” Patient models are being used to predict and discover novel sensitive and specific risk biomarkers, predict disease progression, design optimal treatments, and discover novel drug targets. Applications include cardiovascular and neurological diseases, cancer, and critical care and emergency medicine (1).

Rai L Winslow, Director Institute for Computational Medicine, The Raj & Neera Singh Professor of Biomedical Engineering, The Johns Hopkins University

How is computational medicine changing critical care?

5 Questions

5 Answers

Why Sepsis

What was the starting point for your work on sepsis and septic shock in adults?

A starting point for my work on sepsis and septic shock was reading a paper that demonstrated how every hour of delayed treatment in patients with septic shock could lead to an eight percent increase in mortality, per hour. That statement really stood out because what it told me was the natural time course of evolution of the disease, and whatever was happening in septic shock, was happening very quickly. Because of this rapid disease progression, this suggested that accurate prediction of those patients with sepsis who would progress to septic shock must be based on data collected from the patient on a time scale of minutes rather than hours. The challenge was that this high-rate data is not routinely collected in hospitals.

Data and algorithms

What live data are the algorithms capturing from patients for studying and understanding sepsis and septic shock?

Today’s electronic health record (EHR) is typically used to store data such as vitals and lab results and clinical observations made at irregular intervals and at low rates. Given the rapid evolution of septic shock, we hypothesized that advanced prediction and early detection of septic shock must be based on data collected at the minute rather than hour time scales. This was the driving interest in developing a novel software platform called PhysioCloud. PhysioCloud captures physiological vital signs data at minute intervals from patient monitors. These data are then stored in a specialized database that is designed to capture large numbers of real-time data streams at high-rate. Data collection also includes waveforms, such as ECG, respiratory rates, and SpO2, sampled at 125 times per second. Nowhere else in the USA that I am aware of, is capturing these physiological data from patients, making them a part of the patient electronic health record. Our algorithm uses these high rate data, as well as low-rate data from the patient EHR, to predict those patients with sepsis who will develop septic shock.

The importance of the transition state to septic shock

Computational medicine and algorithms can be uncomfortable terms for medical students, interns and researchers who do not have experience with it. Simply put, how do research and studies such as this help doctors in emergency medicine and critical care units, in managing their patients?

Everyday critical care and emergency medicine physicians ask two questions of every patient they see: what is the state of my patient?; how will their state change over time? The latter is a prediction problem of the sort that data scientists often confront. In the context of sepsis, the physician would like to know if their patient will at some future time develop septic shock, or will their condition improve. If an algorithm can reliably predict those patients with sepsis who will develop septic shock at some future time point, then physicians will have a window of time in which they can intervene to prevent this transition from happening. Our goal was to develop such an algorithm. To do this, we utilized the obvious fact that if a patient has sepsis and their condition is getting worse and possibly evolving towards septic shock, it means their physiology must be changing over time as they get sicker. We, therefore, decided to develop a “risk score,” a number ranging between 0 and 1 that is the probability that a patient will develop septic shock. This risk score was computed in an optimal way from the minute by minute physiological vital signs data complemented by clinical data from the EHR. If this risk score exceeds a threshold value, then we decide that this patient with sepsis will develop septic shock at some future time point. This approach works very reliably, achieving high sensitivity and specificity. It’s the worlds simplest machine learning method. Predicting the transition from sepsis to septic shock can enable physicians the ability to follow their patients and see how various states are evolving over time, so that they can intervene to deliver earlier care. Right now, this approach is being applied in retrospective studies using patient data. In the future, we plan to compute this risk score in real-time, generating alerts for caregivers when the risk score exceeds threshold signaling that patients are likely to go into septic shock.

Pre-Shock

In a recent publication in Scientific Report (2), the new concept of a pre-shock state was outlined. How was this possible to do?

Our work hypothesized that it was possible to identify the presence of a physiological signature in sepsis patients before the clinical onset of septic shock was diagnosed. We were able to identify a signature to calculate a risk score for the pre-shock state. The changes in variables such as lactate and heart rate are so small; they are still statistically significant, but so small. When discussed with physicians, some say that they would not have noticed it. These variables are changing together in a small way, but the algorithm is able to catch the changes together and compute it into a risk score and make useful predictions. Some of our very new work not published yet shows that post-threshold, changes in patient risk score happen very quickly (30-60 minutes) and are very large. We have shown that the larger the post-threshold risk score, the more reliable is our prediction that the patient will go into shock. Positive predictive value can be as high as 80-90%.

Fluids and Vasopressors

Evidence-based studies and protocols such as the SOFA score (3), Surviving Sepsis Campaigns (4) are listed on the American College of Emergency Physician (ACEP) website (5) as well as the SALT-ED (6) and SMART (7) trials. These are referred to by emergency physicians in the emergency department, and EM residents are trained with these resources. How do these studies tie into computational medicine, machine learning and predictive analysis for developing septic shock?

Our algorithm looked at tens of thousands of patients, and computationally phenotyped them through every minute of data using the international consensus definition of septic shock, and based on early warning times, found clinical ground truth. We also discovered that the Sepsis 2 definition had a property that was temporarily unstable. This is to say that the state of a patient with sepsis as defined by Sepsis 2, was changing all the time, and it was not possible to predict ground truth. With found the Sepsis 3 definitions to be temporarily stable with few state transitions. The major factor was that the criteria in Sepsis 2 had included a diagnosis of SIRS before sepsis was considered as a diagnosis, and it was removed from 3. We believe that SIRS was causing frequent state changes, as an ambiguous diagnosis.

We are able to predict those patients with sepsis who will transition to shock many hours before they go into shock. We are also able to identify distinct temporal patterns of the risk score corresponding to patient populations with high (up to 60%) versus low (10-20%) mortality. For each of these groups, we looked at comorbidities, diagnoses such as kidney failure and cancer, but we do not know what the relationship is or what is different about these patient groups and the fact that they are in the 60% mortality pool. We know their physiology is saying they are in the mortality pool, but not why. What this means is how these patients are being treated could be the issue (physicians with different levels of training, and other factors involved in treatment decisions). In our work, patients were classified into high and low risk. We found that patients in the low risk received vasopressors and adequate fluid resuscitation and for patients in the high-risk pool, fewer had received vasopressors or fluids. The question is, why are these patients not getting these things. Our algorithm to predict the transition to septic shock can positively influence treatment decisions made by many physicians, to confirm the value of treatment and prevent the development of septic shock. We’ve also identified and know the time to look for proteomic and genomic biomarkers for the early predictive shock signature that could correlate with this high risk/these measures are not routinely done clinically, and this line of work could be very helpful in understanding the fundamental biology of the very rapid change in patient state when they cross the risk score threshold.

Thank you to Professor Winslow for taking the time to discuss the research involved in computational medicine and investigating the transition from sepsis to septic shock. In closing, regardless of medical specialty interests, medical students around the globe interested in taking a gap year to gain research skills will find the experience invaluable and will be introduced to new ways of thinking, writing, and understanding the scientific influences on patient management and health care. Research such as this in the USA can also be implemented at international hospitals and remote clinics, to further aid patient care and management. There are many areas of interest in which research is taking place in critical care units and emergency departments, and discovering the technology involved such as machine learning and computational medicine, is a step towards understanding the potential advances in the future of medicine and patient care.

Please feel free to share your own particular research area(s) of interest and pose any questions you may have in the comments section below.

References and Further Reading

  1. The Institute for Computational Medicine (ICM) –  https://icm.jhu.edu/
  2. Liu R, Greenstein JL, Granite SJ, Fackler JC, Bembea MM, Sarma SV, Winslow RL. Data-driven discovery of a novel sepsis pre-shock state predicts impending septic shock in the ICU. Scientific reports. 2019 Apr 16;9(1):6145. – https://www.nature.com/articles/s41598-019-42637-5.pdf
  3. Faust J. No SIRS; quick SOFA instead. Annals of Emergency Medicine. 2016 May 1;67(5). – https://www.annemergmed.com/article/S0196-0644(16)00216-X/pdf
  4. Surviving Sepsis Campaign (SSC) – http://www.survivingsepsis.org/Pages/default.aspx
  5. ACEP Statement on SSC Hour-1 Bundle – https://www.acep.org/by-medical-focus/sepsis/
  6. Self WH, Semler MW, Wanderer JP, Wang L, Byrne DW, Collins SP, Slovis CM, Lindsell CJ, Ehrenfeld JM, Siew ED, Shaw AD. Balanced crystalloids versus saline in noncritically ill adults. New England Journal of Medicine. 2018 Mar 1;378(9):819-28. – https://www.nejm.org/doi/full/10.1056/NEJMoa1711586
  7. Semler MW, Self WH, Wanderer JP, Ehrenfeld JM, Wang L, Byrne DW, Stollings JL, Kumar AB, Hughes CG, Hernandez A, Guillamondegui OD. Balanced crystalloids versus saline in critically ill adults. New England Journal of Medicine. 2018 Mar 1;378(9):829-39. –  https://www.nejm.org/doi/full/10.1056/NEJMoa1711584
Cite this article as: Bryn Dhir, "The Research of Predicting Septic Shock," in International Emergency Medicine Education Project, August 12, 2019, https://iem-student.org/2019/08/12/the-research-of-predicting-septic-shock-how-computational-medicine-is-changing-critical-care-in-5-questions/, date accessed: December 8, 2019

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.