Penn Medicine using predictive model to anticipate ER visits by lung cancer patients

Researchers are leveraging data in EHR to identify those at risk, providing alerts to physicians in real time.

____simple_html_dom__voku__html_wrapper____>Using a predictive model that leverages electronic health record data, researchers at the University of Pennsylvania’s Perelman School of Medicine were able to anticipate in a pilot program which lung cancer patients were at risk for treatment in the emergency department.

During a two-week trial period, Penn researchers successfully predicted a third of all ED visits and were able to identify which lung cancer patients were at high risk and low risk for requiring this kind of care. Results of the pilot were presented at the 2017 American Society of Therapeutic Radiation Oncology annual meeting held this week in San Diego.

According to Abigail Berman, MD, an assistant professor of radiation oncology at Penn and the associate director of the Penn Center for Precision Medicine, lung cancer is the most common diagnosis among cancer patients who visit EDs because of infection, pain or other symptoms. In fact, about 40 percent of lung cancer patients go to the ED during the course of their treatment, of which 60 percent of the visits result in hospital admissions.

“We realize as clinicians that our lung cancer patients wind up in the emergency room very frequently, and even though we know our patients well, we don’t have a great way of knowing who is at risk for going to the ER,” says Berman. “Obviously, we want to take care of our patients before they have to go to the ER and keep them out of the hospital whenever it’s possible.”

Also See: Johns Hopkins develops electronic triage tool to better prioritize ED patients

The predictive model identifies key comorbidities such as cardiac arrhythmia, hypertension and liver disease, as well as flags specific symptoms, including nausea and vomiting. In addition, the model factors in lab results like abnormal platelet count, creatinine and white blood cell count.

“Our model pulls all of this together and weighs each factor to determine a personalized risk for each patient at any given point in time,” adds Berman. “It also gives physicians real-time alerts when a patient is deemed to be at high risk.”

Initially, the model was developed with data from about 2,500 patients and then was validated with a smaller set of data. In a two-week pilot, the model anticipated 68 of the 207 ED visits (33 percent) by lung cancer patients and also demonstrated potential in categorizing patients into risk levels. For example, 13 of the 131 patients identified as high risk (10 percent) actually presented to the ED. And only 10 of the 678 patients categorized as low risk (1.5 percent) had an ED visit.

Overall, Berman notes that when it came to differentiating between high- and low-risk lung cancer patients, those in the pilot who were deemed by the model to be high risk were 6.6 times more likely to visit the ED versus those seen as low risk.

Going forward, Penn researchers want to be able to categorize the reasons for each ED visit and the actions taken by clinicians during the pilot. In particular, they plan to leverage natural language processing to improve the model’s predictive value.

“You can get a lot of information from an electronic health record, but a lot of the best information is actually embedded within clinician notes,” observes Berman. “One of the best approaches, in general, is to try and dig through those clinician notes as well—the best way to do that in an automated fashion is natural language processing.”

Ultimately, she says the goal is to develop a tool for early intervention that will enable clinicians to help patients avoid ED visits by reaching out to them preemptively and scheduling an outpatient visit, better managing symptoms, and other proactive steps.

“Right now, the model only applies to lung cancer—because that’s what it was validated for—but we do hope to expand it to other malignancies as well,” Berman concludes.

More for you

Loading data for hdm_tax_topic #better-outcomes...