Looking to slash your readmission rates using big data but not sure where to start? It's best to hear the stories from the folks who have done it successfully. UPMC's analytics team is one of groups ahead of the curve.

Pamela Peele, the chief analytics officer for UPMC's insurance division, together with her team of 25, have done what many hospitals and payers are just beginning to do: They developed a conditional readmission model on the payer side that delivers a readmission risk prediction score before the patient even walks through the door, and then blending it with a provider-side model.

[Learn more: Meet the speakers at the Patient Engagement Summit.] 

"So we're predicting your readmission rate, and you haven't even been admitted yet," Peele told Healthcare IT News. "And we're pushing that information over to our provider side, so when somebody presents and they're being admitted, the provider can see the a priori readmission risk that we've already calculated and can act upon that risk starting at the point of admission."

Peele, who will be detailing UPMC's case study at the HIMSS Media Big Data & Healthcare Analytics Forum this November in Boston, said this model runs off claims data and is done prior to readmission.

Then there's a second model the hospital runs, which extracts data from the electronic medical record and is updated every 15 minutes. This model predominately uses lab and clinical metrics and runs during the admissions process.

Bringing two different approaches to big data to the table and integrating that information into their operational flow proves to be the hardest part, said Peele. Not only that, but also resolving those discrepancies between the models.

[See also: Big data 'not for the weak' and No interoperability? Goodbye big data.]

"Both of those models, they perform almost identically," said Peele. But at the point of discharge, what happens when these two models don't agree? "So coming into the hospital you look like you would be a low risk for readmission, but going out of the hospital, the hospital model says you're a high risk, or vice versa," she added. What happens then?

Two words: big data.

"That's where we bring our big data to play once again, along with unsupervised machine learning to come up with a set of rules," said Peele.

Hospitals, she continued, need to be able to work off a check list, rather than wait for some time-consuming data application running in the background.

Peele and her team used unsupervised machine learning to create the rule set that dictates which model leads when the two models disagree.

For those looking for more concrete specifics, Peele will be highlighting her team's approach and strategy in Boston Nov. 5, in her session "Using Big Data and Machine Learning to Reduce Readmissions". She'll also share how they used their data to better understand optimal timing for follow ups after patient discharge.

Then there's the integral piece of, once you know that optimal timing, then monitoring physician performance relating to their optimal timing of follow ups, post discharge, which UPMC has started to do.

The overall readmission results for certain patients, as Peele pointed out, speak for themselves. "We have hard numbers around that," she said. But she didn't want to give it all away before the big presentation.