The Big Promise of Predictive Analytics in Healthcare

At Duke University’s Fifth Annual Technology and Healthcare Conference, Eric Siegel, founder of Predictive Analytics World and executive editor of the Predictive Analytics Times called new predictive analytical tools “inevitable” disruptions to the way physicians make treatment decisions and patients receive care.

Whether you’re at a casino in Las Vegas, or a patient on the active arm of a clinical trial, no knowledge is more coveted than what’s going to happen next.

Of course, no one can know with certainty what the future holds – there are far too many variables, known and unknown – but that’s not really the goal of predictive analytics anyway. For his purposes, Eric Siegel defined predictive analytics for conference attendees as “technology that learns from experience – i.e. data – to predict the outcome or behavior of individuals.” But even that definition is a bit deceptive; technology itself is subject to the same chaotic undercurrent that defines the lives of human beings and their machines.

The famous baseball statistician Bill James, who brought scientific analysis and big data to bear on the sport back in the 1970s, began his project by obsessively studying box scores in an attempt to understand why some teams win and others lose. Despite James’s undying interest in hard numbers and percentages as tools for understanding and predicting the game, he always stressed the anomalous factors, and the need to wed traditional player statistics with the more ethereal characteristics the players embody. Things like luck, the effects of playing at home or away, and clutch performances in the bottom of the 9th, with two outs and the bases loaded, turn out to be pretty unpredictable.

This isn’t an attempt to debunk predictive analysis as a marketing tool and a potential route to better health outcomes. The ROIs are written on the walls. But the dramatic increase in the number of people wearing biometric sensors, paired with all of the “listening” or spying campaigns being conducted on social media platforms, to name just two small streams in the flood of new and accessible data, have made certain commercial enterprises increasingly confident about the degree to which they can predict an individual’s behavior.

That capability, always described at conferences as “the holy grail” or, in Siegel’s parlance, “the golden egg,” is starting to make the question of what technology can accurately predict about people less interesting than what it still can’t.

At any rate, Siegel got around to admitting that predictive analysis is “not necessarily [about] predicting individual outcomes,” but is more about segmenting risk levels. The easiest and most basic form of predictive analysis begins with a decision tree. But even before constructing the decision tree, the crucial first step is to prepare the data by organizing it so that two time frames are juxtaposed: historic data on the one hand, and present day data, which companies would like to be able to predict. Siegel says the relationship between past data and present data is analogous to the relationship between present data and future data. Once the data is prepped, the decision tree can take root.

In an example from Chase Bank’s mortgage business, Siegel described the top of the decision tree as an interest rate of <7.94%. By asking a series of yes or no questions, involving income level, total mortgage amount, lone-to-value ratio, etc. etc., Chase was able to very accurately predict an individual’s risk of loan defection.

In healthcare, the idea is that a similar decision tree, based on extensive patient data and clinical drug information might help bring personalized medicine a lot closer to home for many patients. And it might also upend traditional treatment pathways and protocols, since no two people are exactly alike. Siegel said predictive analytics at the patient bedside is “inevitable,” although it could start happening in five years or 20. Not because the technology and methodology isn’t ready for prime time, and not because predictive analysis is too complicated, but because “cultural change is hard…we have to learn to trust the machine.”

The three most promising applications for predictive analytics in the healthcare space, according to Siegel, are in the areas of clinical (diagnosis, outcome prediction, and treatment decision-making); marketing; and insurance coverage. In his presentation, Siegel cited examples of pharma companies who have dabbled in clinical predictive analysis – GSK has experimented with predicting clinical trial enrollment, Pfizer with predicting health outcomes – but Siegel himself hasn’t fully waded into the healthcare industry as of yet. That will change next year; Siegel announced an inaugural healthcare-focused conference that his organization, Predictive Analytics World, will host in Boston next October.

Prediction has come a long way since Nostradamus. Today’s predictive analysis isn’t concerned with causality, for two major reasons. One, it’s often impossible to determine; and two, it’s largely irrelevant. What matters are the correlations, which readily emerge once the datasets grow large enough. The owners of those datasets, or the people and machines that have the best access to them, are in a position of power that will only increase. Toward the beginning of his keynote, Siegel told attendees “your experience today depends on how organizations and companies treat you.” The most unsettling thing about that statement is that it’s probably true.