Predictive Medical Technologies claims that it can use real-time, intensive care unit (ICU) monitoring data to predict clinical events like cardiac arrest up to 24 hours ahead of time. Effectively, the startup’s algorithms are new types of medical tests that an ICU doctor can take into consideration when deciding on a course of treatment.
Predictive Medical Systems is based in the University of Utah’s medical accelerator, which is attached to a hospital. The system will soon be tested on a trial basis with real patients and ICU physicians.
I recently talked to CEO Bryan Hughes about using data in diagnosis. Our interview follows
What kinds of data is already available from hospital electronic medical records (EMR) and patient monitoring systems?
Bryan Hughes: We require that a hospital be at a certain technological level, in particular that the hospital has an EMR solution that is at minimum classified as Stage 4, or a Computerized Physician Order Entry system. Only about 100 hospitals in the U.S. are at this stage right now.
Once a hospital has achieved this stage, we can integrate with their computer systems and extract the raw data coming from the monitors, lab reports and even nursing notes. We can then perform realtime patient data mining and data analytics.
Our system works behind the scenes constantly analyzing the raw patient data coming in from a variety of sources like chemistry panels, urinalysis, micro biology, respiratory and bedside monitors. We attempt to alert the doctor early of an adverse event such as cardiac arrest, or that a patient might be trending toward an arrhythmia or pneumonia.
How does the system integrate into an ICU doctor’s existing routine?
Bryan Hughes: Depending on the technological development of a hospital, doctors either do their rounds in the ICU using a piece of paper or using a bedside computer terminal. Older systems might employ a COW (Computer On Wheels).
For hospitals that are still paper based, they have to first get to the EMR stage.
It is surprising that health care, the largest and quintessential information-based industry, has failed to harness modern information exchange for so long. The oral tradition and handwritten manuscripts remain prevalent throughout most of the sector.
For hospitals that have an EMR, there are still several fundamental problems. The single most daunting problem facing modern doctors is the overwhelming amount of data. Unfortunately, especially with the growing adoption of electronic medical records, this information is disparate and not immediately available. The ability for a clinician to practice medicine is rooted in the ability to make sound decisions on reliable information.
Disparate information makes it hard to connect the dots. Massive amounts of disparate information turns the dots into a confusing sea of blobs. The dots must be connected in a manner that allows the doctor to make immediate and intelligent decisions.
We look at the current trends and progressions of disease states in the now, and then look at what may be happening in the next 24 hours. We then push this information to a mobile device such as an iPad allowing the doctor to see the clinically relevant dots, allowing them to make better decisions in a timely manner.
Eventually we hope to expand to the entire hospital. But for now, the ICU is a big enough problem and a great starting point.
How do you use data to predict outcomes like cardiac arrest?
Bryan Hughes: We have two first-generation models: cardiac arrest and respiratory failure. We plan to apply our novel techniques to modeling sepsis, renal failure and re-intubation risk.
Without giving away too much of our secret sauce, we use non-hypothesis machine learning techniques, which have proven very promising so far. This approach allows us to eliminate any human “expert” bias from the models. The key then is to ensure that the data we use for development and training is clean. It is only now that medical data is in electronic and structured form that this is becoming readily available.
What kinds of data mining techniques do you use in the product?
Bryan Hughes: We use a variety of techniques. Again, without giving too much away, our approach is to use transparent algorithms rather than a black box approach. We have a patent strategy that allows us to effectively place a white fence around our technology while allowing the academic and medical community to review our results.
How do you judge the accuracy of the algorithms?
Bryan Hughes: To date, our results have been proven using retrospective models (historical ICU monitoring data and outcomes). Our next step is to deploy our technology into a validation trial — a validation trial produces evidence that a test or treatment produces a clinical benefit. That trial is about to start at the University of Utah Medical Center in Salt Lake City.
Once the integration is completed in the next several weeks, we will be running a double-blind, prospective study with patient data. While this is only a validation trial, we are following the FDA guidance. Once the trial is up and running, we plan on expanding the validation trial to include several more hospitals. It will be at least 12 months before we start any formal FDA trial.
How is the system updated over time?
Bryan Hughes; We have developed a unique architecture that allows the system to reduce the experiment to validation cycle to 8 to 10 months. Typically in the medical community, a hypothesis is developed, a model is built and then tested and if valid, a paper is published for peer review. Once the model is accepted, it can have a life span of several years of adoption and application, which is bad because as we know, information and knowledge changes as we learn and understand more. Models need to be consistently re-evaluated and re-examined.
Are any similar systems available?
Bryan Hughes: None in the ICU, or even dealing with patient care, that we have found to date. In other industries, predictive analysis and modeling are pretty common place. Even your spam filter employs many of the techniques that the most sophisticated risk analysis system might use.
Photo: ekg by krzakptak, on Flickr