InetSoft Webinar: Compress the Time Between Analytics Question and Answer
Below is the continuation of the transcript of a Webinar hosted by InetSoft in May 2018 on the topic of Machine Learning in Healthcare. The presenter is Abhishek Gupta, Product Manager at InetSoft, and the guest is Jim Reynolds, CTO at Health Analytica.
Abhishek: Right, well suffice to say it certainly sounds like you've been able to compress the time between analytics question and answer. This isn't a long term batch issue, and that allows investigation to take place where one answer to a question leads to the next question, and you can't get there rapidly without that process. So tell me a little bit about speed, not just speed of data but speed of investigation. How is your data science solution supporting that?
Jim: So one of the most difficult things to do it in investigation is if you imagine a dataset of a billion services, and that's a pretty good number of services. What I mean by service is when you go to the doctor, and they do a procedure, that's typically called a service, and when you have giant dataset of a billion or more of these services your job is to go figure out today which ten are bad.
That's a classic needle on a haystack. So what we've done is provided a risk management framework over the top of that where we have advance sensors in our data science platform that go and look for all these combinations of bad behaviors up in mass, and the analytics will run these things very, very fast, and they supply up all of the services which are highly questionable.
When we take the services that are highly questionable, and we run them through our risk framework to allow our calculation of what providers are engaging in the riskiest behaviors then how do they rank relative to other providers who maybe doing the same kinds of behaviors.
So from a user perspective I can go from a bucket of high risk providers, drill onto their metrics look at how they compare it to other providers and then drill all the way down into the exact services that shouldn't have been happening and be able to quantify what they are in a time series fashion, and we allow the user to do that in four clicks.
That's a significant time savings for an investigator because in the grand scheme, what typically happens is they just get a list of bad claims, and then they have to go figure out why are they bad and how am I going to go combat it. But we supply all this for them, and it just keeps crunching on these things on a continuous basis.
Abhishek: Right, how long, Jim, have you've been using the column store architecture behind HPE Vertica, and what did you use before that?
Jim: So we went on a journey like probably a lot of other data store technologies. We tried out a whole bunch of things, and originally when the problems were smaller a database like Postgres or some other open source MySQL, pick your favorite database, those were fine until we started reaching a certain type of scale and a certain type of query.
Then the challenges really started hitting people, and so we went on a journey of looking at the traditional column stores prior to things like Vertica and Cassandra. We looked at other graph databases. We looked at other massively parallel databases, and as we started moving through them. Either they were really, really expensive to maintain, really hard to set up, or very difficult to make highly available.
As we started building that criteria out and swapping out the technology, we found Vertica, and we were able to satisfy the most number of requirements from our scale and speed up, but we were also able to do it with a very small number of nodes in comparison to other database technologies, and that was a big deal for us.
We did not have to build out a giant multi hundred node data store to get the performance that we needed, and for us that was a big deal. In addition to that Vertica is extremely manageable, easy to manage technology, and those were all things that we really liked, and it has been for us an extremely cost effective store for our column store.