Using Predictive Analytics to Ensure Coding Accuracy and Maximize Charge Capture

Predictive AnalyticsBy Sarianne Gruber
Twitter: @subtleimpact

With third party fee schedules diminishing, health providers face the pressure of an economy that is making it very difficult to collect fees and very important to identify charges for every legitimate service. Electronic Health Records and Hospital Information Systems with their large data warehouses have fueled revenue cycle management to improve provider productivity. Unlike many other industries, employing their data and using analytics for some time now, hospitals and physicians are finally being “primed for disruption”. Financial services, telecom, online shopping sites and social media companies continuously model our behaviors for credit scores, cross-selling products and networking, and they grow profits with big data and predictive analytics. Our webinar “Using Predictive Analytics to Ensure Coding Accuracy & Maximize Charge Capture “ delivers an in-depth and informative discussion on why charge capture and correct coding is a challenge and how statistical and data science tools can put health providers back on top of revenue recovery. Our presenters are Dan Ward, Vice President of Strategy and Paul Bradley, Ph.D., Chief Data Scientist from ZirMed, Inc.

It is all about getting the correct payment for the service provided, getting paid promptly and eliminating delays in payment. Yet, missing charges, coding assignment errors and overcharges has historically rivaled the best attempts at charge integrity. When the identification of missing charges took on a monetary impact, the importance of identifying anything that compromises the integrity of the claim became a priority. Dan points out the “evolution of charge integrity” starts with manual audits, followed by bill editors and bill analyzers in the early 2000s, and then uniform payer logic roots rules-based software. Dan shares how changes in today’s clinical environment no longer supports rules base coding, simply stated “if A then input A”, rather mandates advanced analytics that draws from patterns and correlations. These models have a much higher likelihood to detect if a charge is missed (referred to as “leakage”), a claim will be denied or if a patient will pay a portion of a bill.   Basically, data mining and artificial intelligence are anomaly detection tools that leverage data in ways to identify any deviations from an expected norm. Dan contends artificial intelligence takes the best of all manual audits and puts it in a data mining procedure that runs in less than five hours and produces 300,000 to 400,000 iterations of connections to identify charges and trends.

“We are looking at 5,000 to 9,000 data elements for a given patient visit”,  affirms Chief Data Scientist Paul Bradley. Hospital data comes from disparate sources such as claims, registration, clinical, payer and third party sources. All the information is integrated, aggregated and validated before the modeling process begins. A training data is created with twelve months of historic data, and contains various elements such as type of visit, admit and discharge dates, diagnosis and procedure codes, admit source, charges and reimbursements, and prescriptions and labs. Models are built depending on the event of interest e.g., a missing charge, a denied claim, or the propensity to pay. Applying predictive modeling algorithms creates correlations between patient level and visit elements. To look at different ways the data elements connect, there are decision trees, neural networks, regression, association rules, Naïve Bayes, and clustering. To achieve higher accuracy and more stable models, Paul recommends consensus prediction, a technique that combines two or three (or more) model types where predictions agree. As an example, a medical GI practice with a lot of split cases performs 90% of its routine colonoscopies in the GI lab and special cases in the OR. The vast majority of cases will have 750 revenue codes with appropriate GI charges; the remaining 10% will have OR time only. Building linear and nonlinear models with independent decisions, and then bringing the models together for a unanimous or highly leveraged consensus results in a aggregate model that “learns” this anomaly for OR cases exists. Without employing these methods, a model may flag OR cases for no GI charge, whereas, aggregate models will expand on new data patterns for a more robust model.

Sign on to the webinar and listen to Dan depict how accounts can be missed in a rules based approach and why these cost charges were able to be captured with artificial intelligence and data mining. Learn more about predictive analytics and modeling methods with Paul, sign on to our webinar.