High-Performance Scoring of Healthcare Data

5 Min Read

Copyright © 2009 James Taylor. Visit the original article at High-Performance Scoring of Healthcare Data.Syndicated from Smart Data Collective
Natasha Balac from UC San Diego and Michael Zeller from Zementis (their product was blogged here and their support for the amazon.com compute cloud was discussed here) presented on the use of Medicare and Medicaid data to […]


Copyright © 2009 James Taylor. Visit the original article at High-Performance Scoring of Healthcare Data.

Syndicated from Smart Data Collective

Natasha Balac from UC San Diego and Michael Zeller from Zementis (their product was blogged here and their support for the amazon.com compute cloud was discussed here) presented on the use of Medicare and Medicaid data to detect and prevent fraud. The high computing center at UC San Diego (San Diego Supercomputer Center or SDSC) has a huge (3PB+ of disk for instance) data infrastructure and lots of computing resources. They keep a lot of data and do all sorts of data mining on that data working on a variety of public and private projects.

One of these is for the Centers for Medicare and Medicaid Services (CMS) which has been chartered to detect and eliminate Medicare/Medicaid provider fraud. CMS is bringing together the data from all 50 states and all sorts of stakeholders to try and find fraud. The project handles 30TB of data (all HIPAA and FISMA regulations among others) and SDSC is providing a platform for this project. The project has 50 states, lots of flavors of fraud. The claims data is being mined using various rules for known-fraud (several states had these already) but fraud changes all the time so also need predictive analytics to find new and unknown kinds of fraud. SDSC focus on profiling and detecting fraud (and errors) using data from insurance claims, pharmacy, doctors and more. This project needed not only to handle large amounts of data and lots of
transactions but also do so flexibly.

The project uses R to build regression trees and neural networks and then exports these models using PMML (Predictive Model Markup Language – an XML syntax for predictive models) to the ADAPA engine from Zementis. Zementis ADAPA is a decision engine that supports rules and analytic models and is very focused on open standards like PMML. They use an open source rules engine (JBoss Drools), support JSR73 for data mining and deploy as a standard decision service so that any system can call the decision service and get questions answered.

Adding R to build the models allowed open source and standards-based development of the models also. R is an integrated suite of software for data manipulation, calculation, visualization and data mining. Zementis has been contributing code to this project so it can support PMML export and lots of other companies integrate with and contribute to the R project.

PMML is not as well known or widely used as it should be. It is a robust and usable syntax for describing predictive models. It provides a clear separation between model development and model deployment, helping analytic folks focus on building the model and IT people on deploying it. Lots of vendors support PMML (IBM, Oracle, SAP, SAS, SPSS, Fair Isaac, Teradata, Microstrategy, KXEN). PMML also supports the transformations needed for a model. All the pre- and post-processing can also be described in PMML, helping address one of the big issues when models are implemented.

The project found that deployment of models (from R to ADAPA) was almost instant – little or no delay. The models were built against 300,000-500,000 rows and showed 90% accuracy or so. The project found that this approach was scalable and fast while delivering the flexibility they wanted to support the different states and their rules. They are also scoring data as part of their Extract-Transform-Load (ETL) process allowing them to detect fraud before they load into the database. The use of PMML also allowed them flexibility of mixing commercial and open source products.


Link to original post

Share This Article
Exit mobile version