Let’s talk about selection criteria for that AI\ML platform you’re considering. (ed: yes, I know, I said let’s stop calling it AI\ML, ships turn slow)
We’ve been seeing a lot more ads about different AI platforms out there and they’re typically along the lines of “…crunch through billions of rows of data in 9 seconds flat!” – or some such. While those are great metrics to consider, they really need to be second, or even third, on the list of what you’re thinking of.
Here’s how we approach ML platform selection with our customers.
First – you really want to consider how the platform ingests all the data you have. It used to be there was a lot (a LOT) of data normalization – and that shouldn’t be the case anymore. Make sure you understand what data formats you can export (or which and how you will allow connections into your sources) and the capabilities of the platform to ingest, understand and draw inferences from that data.
Second – what questions are you trying to answer? It’s great to just look forward to being delightfully surprised by unexpected pleasant surprise correlations – or causations – and be able to tell between the two. It’s even more important you set out with a list of questions you want optimal answers to and find those pertinent patterns in your data.
Third – accuracy! Even after training the model, is accuracy being actively, reiteratively measured and are you ensuring you keep training your model as you discover new, pertinent, patterns? This is critically important to ensure you’re tracking to reality and new emerging and changing old patterns. Data, business – the world – is never a static snapshot in time. This, by the way, is where a lot of ML projects come up short – the process of ensuring reiterative training doesn’t occur. ML is not a one shot.
A quick shout out to ClosedLoop.ai. No they didn’t sponsor this post, we just really like their current direction for healthcare!
Finally, keep in mind that most ML platforms today are focused on post-encounter or post-service transactions. If you have pre-encounter or pre-service questions or patterns you are looking for, consider something that’s more of a heuristic platform. We’re still researching and sorting through those now, and tinkering with one of our own making. We’ll make sure to update when we have something tangible.
Thanks for tuning in!