Bayes, Reagan, and Digital Transformation
For a while now, I’ve been on the 7-Minute Workout. This being 2015, of course I use an app to guide me through the routine, complete with coach’s voice, whistle, badges for staying on track, “in-app” purchases – the works! I’ve actually been pretty diligent about sticking with the program: I’ve missed only two days in the month. (Both times, I lost a “heart”, which the program promptly offered to sell back to me for 99 cents, or stock them up “in bulk” for a lower per-unit price!) Having had a fairly active day yesterday, I decided to skip the workout – but I didn’t want to go through the rigmarole of “losing a heart”. So I simply started the program and proceeded to fold laundry; seven minutes and fifty seconds later, the program declared “Workout complete!” and duly recorded my achievement.
What went wrong here? The app clearly has a model of the real world in which tapping the “Begin Workout” screen is a proxy for an individual actually performing the workout, and not tapping it during a day is a monetization opportunity. Unfortunately, this is a pretty weak model and easy enough to fool. And while this example may seem contrived (and to some extent it is), consider the very real problem in e-healthcare of determining a geriatric patient’s adherence to his complex post-op medication protocol at home. How, in this digitally instrumented and interconnected world of ours, do we determine with high confidence whether the patient is taking the right medications at the right times, when he lives by himself and also has the onset of dementia? A major simplification of the problem – by packaging together the right medications for each administration – is being done today by companies such as PillPack and Dispill. But the fact that an individual pack has been torn off, or that a blister pack has been opened, is only a proxy for the eventual ingestion of those medications by the patient. How sure are we that the pills didn’t get forgotten on the counter top, or that they didn’t end up in the garbage can rather than going down the patient’s alimentary canal?
In the good old days of enterprise software, we would simply mandate “Click the DONE button when complete”, train employees to do this, and that would be that. But ours is the world of enterprise digital transformation, where the boundaries separating the enterprise from the customers are fuzzy, and where such an old-school solution would fail to delight the customer and be discarded. And anyway, could you really trust the assumption that an 83-year old arthritic senior citizen would open the app and click the DONE button after taking his medications, reliably, four times a day?
The problem is that most apps or other client-facing software are built upon simplistic and binary assumptions about the real world. Statisticians have long known this to be insufficient, as evidenced in Box’s aphorism that “All models are wrong, but some are useful”. It is time that software acknowledged the messiness of the real world and abandoned notions of certainty. Mathematical proofs of the classical variety rarely happen in the real world; rather, we deal with patterns of uncertainty and degrees of confidence in our beliefs about the world. And there are rarely single unambiguous pieces of evidence that unequivocally establish the truth; rather, we must sort through multiple small pieces of noisy evidence, possibly contradictory or inconsistent, and tease out from this ensemble an approximation to the truth.
Think about that fitness app. How might it have built a more robust model of whether I was actually exercising? Perhaps it could have tied up with my Fitbit (which is already Bluetooth-paired with my phone) to check my heart rate after I started the workout. An increase in heart rate is a good indicator that I am exercising, although it still does not imply that I am actually doing the specific exercises the app is directing me to do.
In the case of the geriatric medication problem, we can all probably agree that the “gold standard” of implanting a nano-tracer within the medication and tracking its presence in the stomach (at a reasonable cost) is likely unachievable for a while. What do we use as a reliable proxy for ingestion that is simultaneously non-intrusive and not susceptible to false alarms? I don’t have an easy answer. I’m sure many in the healthcare and pharmaceutical fields are working on the problem.
Whatever the solutions to problems such as these are, we need to accept that they must be inherently probabilistic in nature. We start with an a priori level of confidence in our worldview; additional observations produce an a posteriori level, which becomes the new prior; and we iterate. In other words, Bayes’ Theorem becomes central for reasoning about an interconnected world deluged with big data from millions of sensors, and essential for separating the signal from the noise. Or, in the words of the Gipper’s signature translation of the Russian proverb Доверяй, но проверяй: “Trust, but verify.”
And yes, I did faithfully complete the 7-minute workout today. (Or at least I’m telling you I did.)
Image Credits: mapmyfitness.com