Need the Right Answers? Start with the Right Questions
Time series waveform data is essential to AI and patient-centered care
I recently attended a medical conference where a physician spoke about how AI could change the face of healthcare. As he took the podium, he explained that the reason he had gone to medical school was to save lives.
“I’ve saved a lot of people,” he said. “But I lose people, too. And the reason I lose them is because, in a single minute, I have to process over 300 data points. Machines are never going to make decisions for me, but I need machines to process the data faster to help me make the best decisions.”
His dilemma is one that’s familiar to all clinicians: Their ability to access all the data they need, as fast as they need it, is hindered. We are moving toward a future where AI could help care providers make critical, life-saving decisions, but the roar over its potential is drowning out conversations we actually need to be having about it. Getting too far ahead of ourselves without asking the right questions is putting the cart before the horse.
We know that a good diagnosis is contingent upon solid information. So let’s start with asking ourselves a good question: What data do we need to build the best models?
Artificial or Augmented?
One of the characteristics that differentiates AI in the healthcare industry is its tremendous human impact. While artificial intelligence is making inroads in every industry from aviation and agriculture to personal finance and transportation, the healthcare space is unique, and must be approached differently.
In healthcare, machine learning is different because the resulting system will not be autonomous. There is never going to be a time when it won’t require a human member of the care team to review the data and then make and implement life-saving decisions.
Rather than “artificial intelligence,” we are actually talking about “augmented intelligence,” a term that paints a much more accurate picture of what we will find on the floor. Machines will never make decisions for doctors, but they’re essential for processing the data to help them make more educated decisions — as fast as possible.
As we race toward making AI and machine learning a part of the exciting future of healthcare, there is a key lesson to keep in mind: The information we are able to get out of AI is only as good as the information we put in. If we want life-saving clinical decision support, then the data we use to inform those decisions is just as critical.
Specifically, we need to make sure that we are gathering the right data, at the right time, within the right context.
EMRs Look Back, Not Forward
Today, the main source of data fueling AI is the electronic medical record (EMR). EMRs give us more information than was ever available before. They have become indispensable when it comes to documentation and reimbursement, and they have come a long way in supporting the care process by providing excellent clinical summaries. Yet now, we are asking them to provide a real-time indicator of patient health and trajectory, something they were never intended to do. Looking back is one thing – predictive analytics is another entirely.
Saddling an EMR with that expectation is simply unrealistic. They just weren’t designed to provide the full patient end-to-end trend data that care providers need to make critical life-and-death decisions. If we want to realize the vision of augmented intelligence in the healthcare environment, we have to begin by getting the right data. 1
You Can’t Beat Beat-to-Beat
“Just listen to your patient; he is telling you the diagnosis.” This medical maxim attributed to Sir William Osler emphasizes the need for communication beyond face value. An important place we find that information is in the beat-to-beat time series waveform data from the biomedical devices monitoring a patient. In a critical care environment, for example, patients may be connected to eight or more monitoring devices, generating more than 800,000 samples of patient data per hour.
Within that data is life-critical information needed to fuel AI models and expedite care and intervention, and improve outcomes. Is it crucial? Yes. Accessible? No. Unfortunately, in today’s healthcare environment, most of the time series waveform data from biomedical devices isn’t stored at all, partially because many of these devices were created when a real, live person sat at the patient’s bedside every minute to read and interpret data. What is stored is siloed and locked down by the thousands of device manufacturers, each with their own proprietary formats and interfaces that don’t talk to each other.
Therefore, the next critical question when we look to enable AI in Healthcare is this: How do we integrate the data from these disparate devices to positively change patient outcomes?
Creating a New Standard of Care
Fusion of time series data waveform data with EMR data is the key to realizing the vision of augmented intelligence to help the healers taking care of patients improve care. Remote access to that data to enable software-based monitoring workflows, as well as integration of this data into the EMR, should be the new standard of care.
Why? Because it is critical that we go beyond reporting a clinical summary that lives in the past and move to presenting a complete patient story that’s happening in real time. Empowering care teams with visual tools rooted in time series data allows them to make a difference. We need to lay the foundation to first give them unlimited access to full patient history and then be able to leverage that historical data for the creation of patient-centric AI. When we do, care teams will have the data they need to save more lives today and create the cornerstone for a new model of care for the future.
How Do We Get There?
Start by asking the right questions. When we work with doctors and other healthcare executives, there are a number of key questions we ask as they consider solutions for time series waveform aggregation and integration:
Can the times series data be embedded within EMR workflows?
Can you access it beyond the bedside and walls of the hospital?
Can you access complete patient history across all connected devices, along with EMR data such as labs and meds?
Can you self-navigate that historical data to build trends across every signal, along with alarms, alarm limits, and other data from the EMR such as labs, meds, and observations?
How far back can you go to see a trend? Critical care patients are sometimes in the hospital for weeks or months – so can you go back that far? How about going back to a prior admission?
Can you send any trend data, including the complete waveform image, automatically to the EMR to expedite orders and improve documentation accuracy and completeness?
Can you send trends to other members of the care team securely to expedite intervention?
Can physicians label the data to begin teaching models within their clinical workflow?
And, can you access the aggregated data easily to build models with data management apps or on your own through open APIs that integrate with standard ML tools such as Apache and Spark?
In my opinion, this is just a handful of the many important questions we should be asking about time series data integration – not only for AI, but also to change the way we practice medicine every day.
Final Thoughts: Don’t Put the Cart Before the Horse
As exciting as the prospect of AI is, we shouldn’t be building these complex puzzles until we have all the pieces. What’s missing from the puzzle is time series waveform data. To realize the vision of AI, we need to ask ourselves the right questions.
Unlocking time series waveform data is the first step. Integrating that data into the EMR and other workflows is second. Providing remote access and tools that enable transformation is third.
Once we achieve all three, we will then be able to flip the current model on its head. We will be able to ensure that care providers are no longer searching for answers in the rearview mirror when seconds matter in real time. We will be able to start looking forward and seeing patterns that can truly change patient outcomes for good. We will be able to realize the vision of AI and create a world of patient-centered care.
1Today, less than 1% of the data generated in the critical care environment actually makes it into the patient’s electronic medical records (EMR), and that information is often incomplete or contains errors.
According to research by the Mayo Clinic, only about 60 pieces of data have the kind of information required for clinicians to make informed decisions — but whether or not they have access to that data is another matter entirely.
Over the next six years, the amount of healthcare data generated is expected to grow faster than any other sector, with an annual compound growth rate of 36% through 2025.