“The term infancy is relative,” says the article's co-author Bin Xie, health services research manager with PCCI, a Dallas-based not-for-profit corporation spun out of the healthcare data analytics work done at Parkland Health & Hospital System.
The decades-old Framingham risk model for cardiovascular events and the APACHE II scoring systems to gauge the acuity of ICU patients are both well known examples of predictive analytics systems, the authors point out.
But very few risk prediction models targeting hospital readmissions had been incorporated into an electronic health record system for easy use and reference, according to a 2011 survey report, published in the Journal of the American Medical Association and cited in the Health Affairs article.
“There are already many implementations across many hospitals in the country and across the world,” Xie adds in an interview. “It could grow into a big, giant adult, so, when we compare it to its potential, it's still in its infancy.”
“We think in five to 10 years, it could really become a big thing in healthcare, especially when we address the difficulty of containing costs and improving the quality of care and the challenge of the growth in the number of senior citizens,” he said.
Just as government penalties for hospital readmissions captured the attention of many early implementers of eHPA efforts, “payment reform is one essential piece to drive this growth” in the future, Xie said.
Predictive analytics has four component parts, according to the authors—acquiring data, validating the risk-prediction model, applying it in a real-world setting and scaling up the model for broader use in a healthcare system. Their article focused on the latter two and the challenges of bringing them to fruition.
Among those challenges are setting up an appropriate oversight mechanism with the right balance between enough control to keep the program operating properly and also affording it enough breathing room to grow and respond to daily events, the authors said. Another is stakeholder engagement, which includes patient consent, particularly when the risk models are still in the early stages of development.
“The first time you go out and experiment, you do need a rigorous framework of the patient's right to know, just as you do in research study,” Xie said.
Other issues that data analytics program planners must address are data quality assurance, patient privacy protections, interoperability of the technology platform and transparency of the risk model.
“Whenever possible, clinicians, in particular, need to be able to 'see into' a risk-prediction model and understand how it arrived at a certain prediction,” the authors advise. Transparency builds needed trust in the model, and it might encourage “crowd sourcing” to improve the model or expand its use to other organizations or settings.
Follow Joseph Conn on Twitter: @MHJConn
