For this project, doctors and data miners are specifically focusing on lung cancer patients. By flagging things like recent lab tests, radiology visits, or patient-reported symptoms, Penn’s team is hoping to come up with a formula that will predict when a patient is likely to end up visiting the emergency room. Right now, the formula can predict an estimated one out of every three ER visits, giving doctors the chance to take action before a patient gets to that point.
“Once we get the alert, we can call the patient ourselves,” said Tracey Evans, MD, an associate professor of Clinical Medicine in Penn’s Abramson Cancer Center and one of the doctors piloting the program. “We can schedule them for a visit to our clinic. We can recommend more frequent follow-ups or increase the steps they are taking for home care. All of this stems from big data, and the hope is it can help keep patients out of the emergency room.”
The term big data gets used a lot these days, but even if you talk to the experts, it’s tough to find a uniform definition. We recently took a stab at the question here, but the most important thing to realize is just how much information the term covers.
“I like the definition of big data that says: It is data that’s so big, it makes you feel uncomfortable,” said Jason Moore, PhD, director of Penn’s Institute for Biomedical Informatics. “You see it, and you’re paralyzed. So we need to help people see the data in a way they can process.”
Just for starters, think of all of a person’s social media posts – the articles they read and share, the photos, and the restaurant check-ins. Together, all of this information paints a picture of who they are. In this case, information contained in a patient’s electronic medical record paints a picture of who they are as a patient and helps doctors understand how to treat them most effectively.
Once doctors can better predict which patients may need urgent treatment, they can either take preventative measures, or direct that patient away from the emergency room and into another facility. With this in mind, Penn recently launched the Oncology Evaluation Center, which now serves as a kind of urgent care facility specifically for cancer patients. Staffed with both oncologists and nurse practitioners, it gives patients the ability to see clinicians who are more attuned to their particular condition when they come in for common complications like infections, fevers, or pain management. The new center is embedded within the Abramson Cancer Center’s regular clinics in the Perelman Center for Advanced Medicine, which also makes it easier to coordinate urgent care needs and regular treatment schedules.
“We hope to be able to help patients avoid being admitted to the hospital,” Evans said. “But even in this case, when admission is required, it’s controlled. This also keeps the patients with doctors they already know, which provides a more comfortable environment.”
The new center can also be on the front lines once the data tells doctors that a patient could be at risk for an emergency room visit. Evans says even if those predictions aren’t perfect, they can still be a tool to help doctors pick up on trends that might otherwise not stand out.
“Good doctoring is pattern recognition,” Evans said. “So to have all of this data here and not use it is ridiculous. It may help us find better ways of doing things that we don’t even know about yet.”
Ironically, the problem with big data in the field of oncology is that it isn’t actually big enough. So much of what we know about the symptoms cancer patients experience and the complications that could lead to an emergency room visit are tied up in physicians’ notes. Even with electronic medical records, the notes themselves are all written in prose, which is hard for a computer to interpret for a project like this.
“There are gaps in our data because of free text data or narrative data,” said Peter Gabriel, MD, MSE, chief oncology informatics officer in the Abramson Cancer Center. “It is very challenging for computers to process this kind of data and turn it into computable facts.”
But even if computers could process this data, it still wouldn’t paint the complete picture. One of the most important pieces of data in the case of predicting emergency room visits is to know which patients have gone to the ER and which symptoms prompted those visits.
“But what if you went to an ER at a different hospital than where you go for your cancer treatments?” Gabriel pointed out. “How would we know? It makes it challenging to get a complete set of data.”
To fill in some of those holes, Penn recently entered into a partnership with Independence Blue Cross.
“By partnering with the insurer, we’re hoping to close the gap on unknown ER visits,” Gabriel said. “That should improve our predictive models for other patients.”
Gabriel also says greater data sharing among medical institutions would be a great step for these kinds of predictions, something he says the National Cancer Institute is pushing for through the Cancer Moonshot Initiative.