Imagine a child observing a beehive for the first time. She would see thousands of bees buzzing in and out, seemingly at random. Even if the beekeeper removed a panel and showed the child the honeycombs and explained the process of making honey, all that buzzing and frenetic activity would make little sense.
Now imagine that the child sits before a powerful computer and asks questions about the bees: how likely is that a bee will thrive so it can build a honeycomb and start making honey? What can we do to help the bee increase her output? The computer is able to identify patterns amid the apparent chaos of ten thousand bees in a hive, and come up with answers to the child’s questions.
A hospital is a lot like a beehive. There are thousands of streams of information that, if properly organized and analyzed, could yield deep insights into the way hospitals work. And that knowledge can be leveraged to produce better outcomes for the patients.
The key to answering questions like these is predictive analytics.
Using predictive analytics to reduce hospital readmissions
Predictive analytics is “the practice of extracting information from existing data sets in order to determine patterns and predict future outcomes and trends”. These data sets are large lists of numbers and documents that make no sense when seen for the first time by a naïve observer, such as our little girl. With the application of tools such as artificial intelligence, machine learning and statistical modeling, predictive analytics can help find order in chaos. It cannot tell the little girl precisely what will happen to an individual bee. Rather, the answers to questions come in the form of probabilities: what are the chances that moving the hive under a tree will help boost honey production? Predictive analytics may say the chances are 36%. The beekeeper may decide to try it.
In a hospital, a nurse manager may use the same sort of analysis to determine the risk of a particular heart failure patient having to return to the hospital within a month. Predictive analytics may return an answer of 30%. With this knowledge, the nurse manager arranges for a visiting nurse to monitor the patient at home during the first month post-discharge.
Early success with predictive analytics
Reducing hospital readmissions for heart failure patients turns out to be perfect scenario for testing predictive analytics. Using data extracted from the Kaiser Permanente electronic medical record system, a study published in 2013 stratified heart failure patients according to risk of readmission. The investigators found that their risk stratification system could save substantial numbers of readmissions by allocating resources such as nursing follow-up.
The results of interventions using predictive analytics are only as good as the design of the models. If the statistical modeling is faulty, the results returned are useless. One famous failure is Google Flu Trends, launched with much fanfare in 2008. Google initially claimed to be able to track flu outbreaks more quickly and more accurately than the Centers for Disease Control. However, subsequent analysis revealed that the Flu Trends algorithm suffered from design flaws that triggered severe criticism from the scientific community. The site was eventually taken down.
For “big data” to reach its full potential, multiple electronic medical record systems must be able to share information. Unfortunately, the several EMR systems do not “talk” to each other, making data sharing nearly impossible. The so-called “interoperability problem” remains a barrier to future success of predictive analytics. Efforts to create bridges between differing systems are ongoing, but agreement on a universal standard EMR language appears to be further off.
As health information technology continues to evolve, and the statistical models are refined, we can expect predictive analytics to provide ever more accurate risk assessments, allowing doctors and hospitals to direct scarce resources toward patients most in need. It might even help beekeepers maintain super-productive hives!