A Few Observations on Big Data, Analytics, etc.

As we never tire of pointing out, health care is full of initiatives which never quite seem to come close to living up to their highly-promoted benefits.  The latest seems to be the collection and use of “big data” and analytics to transform the system.  The latest issue of Health Affairs is devoted to this theme.  (HA Issue)   As you skim through the articles, it is hard not to be struck by the fact that there is almost no reporting of actual concrete results that show any real impact on either cost or quality outcomes.  It is mostly “here are our plans” or “here is what we think we can accomplish”.  Which, aside from our normal skepticism about much-touted approaches, left us wondering why there is so little apparent progress.  Here are a few ideas about what the primary difficulties will be in getting these techniques make any major impact, much less justify their often exorbitant costs.

The first is that much of the data in health organizations is of poor quality and needs significant work to be useful.  This is not like a physics experiment, where everything is exquisitely recorded in a standardized way and readily available for analysis.  And it is not like a grocery store, where you know what was bought, what was paid for it, where the item was located in the grocery store, and if they used a credit card, a lot of variables about the customer.  A lot of important health care data is not amenable to being quantified or objectified.  So the task of getting standardized and accurate data which makes analysis easy and reliable is very, very difficult and expensive.  There is a bit of a chicken and egg problem in addition; we don’t know what data is important until we can do a lot of analysis.  Buying a fancy analytics engine doesn’t solve that problem.  And attempts to force clinicians to record data in buckets of quanta risks losing the essential character of what is being reported.  On the analytics side, there is way too much faith in the ability to break an incredibly complex set of circumstances down into a few clear correlates or cause and effect statements which could be actionable.  Just look at the difficulty in the basic task of a health plan or ACO trying to predict who will be its high-cost members in the coming year.  That task has been attacked for over two decades, and we aren’t much beyond  the state of the art then.   You can spend a lot less money but just saying I will find about 20% to 30% of them by just looking at my high cost people from the prior year and you can find another group by looking at certain chronic diseases which have relatively regular acute episodes, like heart failure.  But right now, there is no clear characteristic that tells you who is going to develop cancer or have that car accident or develop other conditions that make them a high cost case.  Looking at provider practice patterns on the other hand, can be done with existing tools and is more likely to easily yield savings. So I would advise plans and providers to be wary of spending lots of money on tools that don’t yet have a proven value.