Predictive Analytics in Healthcare

Teradata Guided Analytics
Teradata Employee

Our healthcare analytics experience includes various Applications addressing a variety of issues encountered in the complex healthcare delivery network.  Some of the most obvious and pressing use cases are those aimed at improving the quality of healthcare outcomes to provide better levels of patient care.  However, other applications of our analytic tools address different issues surrounding the business of providing healthcare.  One such case recently encountered is a prime example of applying our analytics not only to patient outcomes, but also to business operations supporting healthcare delivery.


Our client is a large, international healthcare provider operating almost 300 hospitals, clinics, and surgery centers.  One of the business challenges they face is the periodic prediction of costs and revenues for each of their hospitals.  Gross under (or over) estimates negatively impact their ability to track operational and financial metrics which materially impact their planning and ultimately their market valuation. 


After digging into their available data, we discovered that the uncertainty originates mostly from just one sub-segment of their patient pool. This segment consists of patients who, at month end, were not yet assigned a Billing Code (also known as a Diagnosis Related Group or DRG in the client’s vernacular).  Assignment of a DRG code significantly narrowed the expected procedure pathway, material costs, length of stay, etc. for any particular patient.  Categorizing a patient into one or more DRG families allows hospital operations to predict a patient’s final cost/revenue with fairly good accuracy. In contrast, patients who were not yet assigned a DRG code were considered completely unknowable.


In reality, any unassigned patient has data that can be used to help predict his or her eventual DRG code. We used information such as medical history, personal profile, and initial symptoms and mined for signals that predict the eventual DRG assignment.  These signals were fed into one or more predictive models to give the best guess of the patient’s expected DRG classification and the expected cost and time metrics.


To validate the usefulness of this approach, Teradata data scientists built and trained a number of predictive models based on SVM, Naïve Bayes, and Decision Trees to determine which had the best predictive accuracy.  Some DRG’s were predicted with 100% accuracy while others proved more elusive. There were at least 2 levels of predictions with different granularities and the best model(s) were ~80% accurate at each prediction.  By using data already available in the client’s systems, we were able to significantly reduce the uncertainty originating from their un-billed patient pool. Unclassified patients who previously caused so much variability no longer wreak havoc on the business operations.


We are happy to employ our analytic techniques such as data mining and predictive modeling to assist clients on any aspect of their business.  However, in many cases (including others presented by this same client) simple reporting and statistical analysis provide sufficient insights without the need for more advanced approaches.  If your prospects have business problems (in Healthcare or otherwise) that might be addressed using our data or analytic capabilities, please get in touch to set up a discussion with our team.