Medical Device Data Can Be a Crystal Ball to Predict Patient State
A keen interest of mine over most of my career has been using medical device data collected at the patient bedside through medical device integration middleware to assist in predicting what was going to happen to that patient state over time. By “patient state”, it is meant the patient’s current vitals signs: a representative measure of patient health.
The data that represent patient state represent or reflect the current values of vital signs and observations of the patient at any given time. This is particularly true of technologically-dependent, unconscious patients in critical care units or in surgery. Aside from thinking of data as a “crystal ball” for predicting the future, the concept of using past information to establish an estimate of future patient state–to predict the future patient state–is a very old concept, and individuals in the aerospace industry familiar with the concepts and application of “multisensor” target tracking and prediction (e.g.: Kalman filtering etal.) should be quite familiar with the approach.
Mathematics to Predict Future Patient State Drawn from Ballistic Missile Tracking
In a way, this is rocket science, or an example application of it. To draw an analogy from the aerospace field, the state of a ballistic or powered object in flight is subject to the equations of motion and its powered flight model. Outside observers (or sensors) viewing the ballistic object can assess its immediate future state based on its current trajectory (i.e., where it’s been, what it’s current patient state is). Using the equations of motion comprising external and internal forces, such as gravitational forces, it is possible to determine within some sphere of confidence where the object will be in the immediate future.
This is an example of the application of “rocket science.” Predicting the future state of a human being is not rocket science–it is more difficult! Yet, certain aspects of this analogy can apply and have been applied successfully to diagnosis and treatment of human conditions. Models of various human systems have been created and the expected behavior or response to such models has been observed. To a large degree, this is the basis behind the diagnosis and treatment of illness using drugs and medical devices. Treatment and diagnostic methods have been derived from observations of the effects on human beings. These observations have been developed through controlled studies, clinical trials, and from experimental observation–even by accident (penicillin, anyone??)
Back to Medical Device Data and Patient State Prediction
This brings me to the use of data derived from medical devices, as part of a medical device integration (MDI) implementation within a hospital environment. These data are primarily observations of patient state, almost telemetry-like in context, from physiologic monitoring and other medical devices employed for the maintenance, diagnosis and treatment of patients. Unlike the ballistic missile analogy, telemetry measurements derived from medical devices cannot be treated as standalone or devoid of specific patient context. For instance, pulse measurements within a wide range of variation are potentially meaningless unless combined with non-numeric type information, such as current patient medical condition, past history, whether the patient smokes tobacco or drinks alcohol to excess, and gender. Yet, these contextual pieces of data when combined with numeric data such as pulse, laboratory and other information, can become quite predictive.
What is There To Predict?
This brings us to the concept of developing alerts (sometimes referred to as “smart alerts”) by combining multiple pieces of information together, passing them through a known model, and determining likelihood of a specific outcome:
“A new paradigm in medical care is the constant surveillance of multiple streams of patient information with the foal of early diagnosis of acute and potentially catastrophic illnesses.”
The ability to predict and assess patient state and assess the likelihood of onset based upon past, recurring and contextual information has been termed “syndromic surveillance” in the literature.
The use of such “syndromic surveillance” lends itself to new, more informative types of alerts and alarms that are not merely uni-dimensional in nature. For example, if a patient’s pulse exceeds or drops below a certain value, then notify the rapid response team (RRT) of an event. By combining multiple sources of data into known models of outcome, the likelihood of events occurring can be evaluated. Such approaches are being studied and have been applied to disease conditions related to onset of sepsis, ventilator acquired pneumonia, and other events that have shown to have high morbidity and mortality rates, especially among the very young and very old.
Medical Device Data Provide A Richer Understanding of Patient State
Medical device data augment the existing electronic medical record system by providing richer, higher-density information that can be updated in seconds and for which certain events are identified rather quickly. For example, changes in cardiopulmonary function which may go undetected in the coarse measurement of hours may reveal critical behaviors over the span of seconds or even minutes. Intensive care patients who are being monitored continuously can fall into this category of patients who are monitored continuously and at relatively high frequency.
Some medical device data measurements need to be combined to provide more telling notification as to the onset of specific conditions. Researchers in several of the referenced articles included at the end of this blog entry have determined that several parameters, when evaluated over time, tend to provide a high level of reliability as to the onset of sepsis hours before the onset actually begins to manifest. Measurements of temperature, heart rate variation, certain laboratory results, and other contextual information have been developed into sophisticated models that, when evaluated together in specific relationship with one another, reveal highly-predictable behaviors and outcomes.
The use of medical device data in support of clinical decision making is still in its infancy. Yet, the possibilities as to use in the clinical setting span far beyond basic clinical charting and post-hoc assessment. As medical device data collection from the bedside becomes more commonplace and the expectations as to the availability of information in real-time grow, new ideas about the use of these data will emerge for patient care management, intervention and prediction. I believe we have just seen the tip of the iceberg.
 Herasevich etal., “Connecting the dots: rule-based decision support systems in the modern EMR era.” J Clin Monit Comput DOI 10.1007/s10877-013-9445-6. 28 February 2013.
 Escobar, G.J., etal., “Early Detection of Impending Physiologic Deterioration Among Patients Who Are Not in Intensive Care: Development of Predictive Models Using Data From an Automated Electronic Medical Record.” Journal of Hospital Medicine. Vol. 7. No. 5. May/June 2012. pp 388-395.
 Sebat, Frank, etal., “A Multidisciplinary Community Hospital Program for Early and Rapid Resuscitation of Shock in Nontrauma Patients.” CHEST / 127 / 5 / May 2005. pp 1729-1743.
 Mayaud, Louis, etal., “Dynamic Data During Hypotensive Episode Improves Mortality Predictions Among Patients With Sepsis and Hypotension.” CCM Journal. April 2013. Volume 41. Number 4. pp 954-962.