Medical Informatics: Predictive modeling enables ACO transition

Discharge practices are “woefully inadequate” and need a makeover in the form of better predictive modeling, in particular as providers transition to an ACO model, said Julie Meek, PhD, clinical associate professor, school of nursing at Indiana University, at the Medical Informatics World Conference on April 8.

“It’s too expensive to be wrong and not catch people,” she said of not taking advantage of harnessing predictive modeling to intervene before a patient’s condition worsens.  

Meek described predictive modeling as a set of tools used to stratify a population according to risk of nearly any outcome and take action before an adverse outcome occurs--improving care and reducing costs. Methods include multiple and logistic regression techniques, and neural networks, an approach to computing in which mathematical structures detect patterns that enable projections in new situations.  

For predictive modeling to have the highest impact, the predictors must be closely associated with your outcome of interest, and be based on a literature review and evidence-based practice findings, she said.

“Using data just because it’s cheap and available is a poor substitute for disciplined inquiry and legitimacy of your choice of potential predictor variables,” Meek said.

Predicting hospital readmission is especially difficult and requires large amount of data. But when looking at 39 predictor variables, Meek said models indicating readmission included a pre-existing disease and past use of high-level care services. However, she said how patients said they were feeling and functioning was the most predictive element of readmission and often providers fail to ask these questions upon discharge.

“No one is asking how we are feeling and functioning,” she said.

Meek recommended that those patients identified as high-risk take a readmission risk profile and be referred to a management clinic where case management and home monitoring is available. The patient and physician would agree on the discharge plans prior to the patient returning home.  

Meek said the development of predictor models requires more nurses and physicians to gain statistical and informatics education.

“We are looking for more advanced-practice nurses,” she said.