Using Predictive Modeling and Public Records in Fraud Detection. Clint Fuhrman National Director Government Healthcare

Transcription

1 Using Predictive Modeling and Public Records in Fraud Detection Clint Fuhrman National Director Government Healthcare

2 Taxpayer Dollars Are Under Attack

3 Opportunities Eliminate the Pay and Chase status quo by looking to other industries, private sector for successful approaches and technologies Identity Proofing/Identity Management Financial Services, Banking Predictive Claims Analytics Property and Casualty Insurance Social Network Analysis Intelligence and Law Enforcement Greater focus on the individuals and entities in the program Are beneficiaries enrolling who they claim to be? Have they disclosed all assets, income, correct state of residence, etc? What are the true backgrounds of the practitioners, officers, agents, etc? What is the risk profile of a provider based on background, associations, etc.? What significant events are occurring between enrollment periods? CMS Center for Program Integrity (CPI) National Fraud Prevention Program focused on prevention and detection that is integrated, risk based, and measurable; four areas of focus: Provider Screening; Predictive Modeling; Data Integration; Case Management

4 The Many Faces of Fraud Over 80% of all suspected fraud cases involve provider fraud. FALSIFICATION OF INFORMATION False coding, altered claims QUESTIONABLE PRACTICES Upcoding, unbundling, cost shifting, prescribing practices, clustering, underutilization, invalid places of service, non contracted providers OVERUTILIZATION Note: Lists are not comprehensive. Medically unnecessary diagnostics, high frequency of office visits, unnecessary durable medical equipment, inappropriate diagnosis procedures

5 Overview of a Data Aggregation/Risk Solution Provider We assess the risks and opportunities associated with people, businesses and assets. Identity Analytics Who are you? Where are you? Who are you related to, and how? How much of a risk do you present? Health Care Background Screening Collections Financial Services Legal Government Insurance Data 34 billion public records 1 million documents added every day 36,000 legal, business, news sources Linking 250M+ unique individuals 1B unique business contacts Analytics Real time analytics Scores to support customer workflow for remote transactions Scores around individual risk/ opportunity Computing 30M transactions/hr <500 millisecond avg search response time ~34 Terabytes in use Health Care Solutions for Commercial Payers 5

6 Utilizing Advanced Technology to Establish Identity and Risk PUBLIC RECORDS PROPRIETARY DATA ENTITY RESOLUTION NEWS ARTICLE LINK ANALYSIS UNSTRUCTURED RECORDS CLUSTERING ANALYSIS STRUCTURED RECORDS COMPLEX ANALYSIS Health Care Solutions for Commercial Payers

7 Claims Analytics Presentation Title

8 Reducing Risk: Advantages of Enterprise Solutions Early detection of fraud, waste and abuse Prioritized results with fewer false positives, which enable more efficient use of investigative resources Alerts concerning adverse changes in the status of individuals or entities accessing benefits or networks Lower claims losses, better cash flow and higher ROI than traditional post payment only methods Consistent control over risk, quality and costs thanks to automated provider screening and monitoring Confidence in knowing that the right providers are being paid for the appropriate services on the appropriate members

9 Analytics: The Value of Tips vs. FWA Software Percent of Respondents 70% 60% 50% 40% 30% 20% 10% 0% Tips Data Analysis Most volume Most savings Health Care Solutions for Commercial Payers

10 Predictive Modeling Adds a Score Plus More Sample Model Score: 985 Significant Edits Plus More Criminal Record Two Sanctions Bankruptcy Copyright 2011 LexisNexis. All rights reserved.

11 Fraud Prevention: Predictive Claims Analytics Internal (Payer) Data Provider Data Diagnosis Data Treatment Data Claim Edits PREDICTIVE MODELING TEXT MINING BUSINESS RULES IDENTITY MATCHING TEXT SEARCH SOCIAL NETWORK ANALYTICS VISUALIZATION FUNCTIONAL COMPONENTS Claims Fraud Identification Provider of Interest Identification DATA SMART ORDER Subrogation Identification External Data Edits Public Records Data Sanctions Data Fee Schedules USER INTERFACE REPORTING ENGINE SCORING ENGINE DATA MART DATA EXCHANGE STRUCTURAL COMPONENTS Social Network Analytics And more Health Care Solutions for Commercial Payers

12 Analytics for Claims Processing Workflow Stopping abusive and fraudulent claims prior to payment will allow customers to devote more resources to providing care to members. Claim edits, usual and customary charges, and network pricing agreements stop only part of the impact of fraud, waste and abuse on healthcare payers Data driven analytics can produce additional claim edits that can significantly supplement the current claims adjudication process Claim level scoring can: enhance identification of claims post pay for audit and potential recoveries be tuned for use in pre pay to stop the most egregious abuses before payment is made Business rules, monitoring for specific treatment codes, and rules for claim routing pre pay or post pay improves workflow Claim Arrives License and sanctions data, criminal history, sexual offender, etc. Claim level edit and scoring results can be supplemented by the identification of providers who consistently bill outside of normal patterns and practice While some providers are relatively easily identified, others exhibit much more subtle patterns that are nonetheless abusive Identifying these more subtle patterns can provide benefit to: In some cases, the SIU In some cases, the claim audit team, and In some cases, the network management team Problem providers can also have their bills returned before payment to have medical records attached Claim Continues in Adjudication Copyright 2011 LexisNexis. All rights reserved.

13 Fraud Prevention: Claim Scoring Using Predictive Models Predictive analytics provides a score for each claim, policy, etc., allowing activity to be concentrated on areas that have the highest probability of financial return Fraud Without Anything Fraud is hidden in a sea of valid claims With Rules Some fraud is captured but much is missed With Predictive Modeling Fraud is concentrated and prioritized for review and mitigation CLAIM NUMBER SUSPICION SCORE Create the target rich environment High Fraud Potential Low

14 Provider Models Models can help identify problem providers early that would not have been identified by other methods Looking at thousands of attributes about a provider or a claim to find a data pattern that makes a robust prediction Models use: Diagnostic codes Treatment codes Provider types Date stamps Identify treatment patterns associated with diagnoses that are characteristic of known problem providers and flag other providers that exhibit similar treatment patterns Copyright 2011 LexisNexis. All rights reserved.

15 Algorithms Supervised vs. Unsupervised Learning Have a specific outcome in historic data Do not have an outcome cluster like together Decision Trees Accurate, conceptually understandable, non linear, non parametric, robust with outliers, missing data, automatic interaction terms Neural Nets Work best with pre transformed smooth data Difficult training time Black Box Regression Most established/widely used algorithm Works well, but doesn t have some of the advantages of trees Works much better on linear data Health Care Solutions for Commercial Payers

16 Social Network Analytics Health Care Solutions for Commercial Payers

17 Challenges Facing Health Care Enterprises Disparate data is spread across separate physical locations Scale of data. BIG Data is getting BIGGER. Adding relationships exponentially expands the size of the BIG Data analytics challenge. LexisNexis has leveraged parallel processing computing platforms and large scale graph analytics for a over a decade. 17

18 Technology advances are enabling a more proactive response The emergence of open-source massive parallelprocessing computing platforms opens new opportunities for enterprises to increase the agility and scale of solutions focused on addressing fraud and abuse. Effectively ingest and integrate massive volumes of disparate data. Process and Analyze exponentially faster than traditional databases. Large Scale Graph analytics, generally thought to be the domain of companies like Google, offer new variables that provide relationship context between events, exposing patterns and outliers that otherwise would be hidden. Can be applied to many other many areas beyond network analysis and social graph analysis, such as epidemiology and mathematics. Suited to revealing well organized fraud networks hidden within BIG Data and generating actionable results. 18

19 Graphic Analysis and Social Network Analytics Graph Analysis Twitter uses Graph Analysis to help the site determine who s connected to whom in the Twittersphere. Google uses Graph Analysis to power its PageRank feature. LexisNexis uses Graph Analysis to resolve Identities and establish relationships Social Network Analysis Graph Analysis that specifically focuses on graphs built on social relationships. 19

20 Trends in Social Network Analytics Addition of External Data Mixes First Party data with Public and Third Party data sources Adds fidelity to existing entities Adds new linkages into the analysis Ads new entities into the analysis Exposes ring leaders and brokers that don t directly participate 20

21 Trends in Social Network Analytics Reliance on Created Data Transform straw into gold Process numerous discrete data points into high value data Advanced Linking Technology Resolve numerous names, addresses, phones, and other info into a Person ID Better accuracy than other resolution techniques Resilient to name, address, and other info changes (i.e. stable over time) Improves detection, simplifies processing, makes results easier to understand 21

22 Targeting fraud using large scale graph analytics Powered by massive parallel processing opensource computing platforms. Graph \Network 3 Billion derived public data relationships between people merged with risk indicators. Graph Analytics examine up to 20 billion data points to create variables that allows for predictive analysis incorporating relationship context and associated risk. Targets fraud across all sectors including Health Care, Financial Services and Government. 22

23 Social Network Analytics On June 6, 2008, the Department of Justice announced the arrest of Felcoranenda Estudillo on charges of defrauding Medicare of approximately $12 million in an elaborate scheme involving home health care services and kickbacks for referrals of patients who were not eligible for services. Estudillo was a registered nurse and operated Wescove Home Health Services from her home in West Covino, CA. Her husband, Oscar Estudillo, owned the business, as well as several others that used the same home address as their base. Mrs. Estudillo is the only person named in the indictment, but records show her husband was the legal owner of the business. The link analysis chart on the following slide was constructed to show the complex array of relationships among Estudillo, her husband, and the varied business they own and operate. Businesses were linked to the Estudillos that were not reflected in the indictment. The identities linked to the Estudillo s in following slide have been masked but are an accurate representation of the relationships revealed by the link analysis. 23

24 Social Network Analytics 24

25 Fraud Detection: Social Network Analytics A top insurer flagged 7 claims as collusion claims Using carrier data alone, we found a connection between 2 of the 7 claims. 25

26 Fraud Prevention: Social Network Analytics Collusion in Louisiana AFTER Advanced Linking Technology is Applied Assigned unique IDs to all parties and HPCC added 2 additional degrees of relative data Family 1 Family 2 Showed 2 family groups interconnected on the 7 original claims plus linked to 11 more. 26

27 Proof of Concept NY Office of Medicaid Inspector General Health Care Solutions for Commercial Payers

28 Purpose of Proof of Concept Applied social network analytics to information provided by the State of New York and public data supplied to identify relationships between a group of New York Medicaid recipients living in high end condominiums located within the same complex and any links those individuals might have to medical facilities or others providing care to New York Medicaid recipients. 28

29 Methodology Derived Public Data Relationships are built from +/ 50 terabyte data base for the entire U.S. population. This is used to build a large scale network map of the Medicaid Recipients and everyone associated within 2 degrees. Patented algorithms used to cluster the network map and generate statistics to measure every cluster. Graph is queried for the clusters with the most significant statistics. For each cluster, if all these recipients are connected.. How many of them are living in expensive residences, owned expensive property or drive expensive cars? How many recipients are contacts of medical businesses? How many medical businesses are associated with any of the people in the cluster? How many are currently receiving benefits? 29

30 City Walk Sample Vehicle Statistics What is the list of preferred expensive vehicles? Make Description # Owned Make Description # Owned Mercedes-Benz 46 Chevrolet 2 Lexus 41 Hummer 2 BMW 27 Jeep 2 Infiniti 13 Nissan 2 Acura 9 Toyota 2 Lincoln 8 Aston Martin 1 Audi 7 Bentley 1 Land Rover 7 Cadillac 1 Porsche 6 GMC 1 Jaguar 5 Honda 1 Mercedes Benz 3 Volkswagen 1 Saab 3 Volvo 1 30

31 Property Deed Reference Counts for City Walk Dominant buyers and sellers at City Walk Name Deeds Held Name Deeds Held Hudson Eight 78 Mike Greem 21 Hudson Five 74 Scott Hill 21 Hudson First 73 Betty Donaway 21 Hudson Nine 65 Al Clark 19 Harry Anderson 45 Dave Miller 17 Hudson Ten 41 Mark Walker 16 Hudson Seven 39 Mike Smith 16 Home Nationwide 33 Val Edwards 15 Hudson Three 33 Eric Garcia 14 Brian Smith 28 Dane Young 14 Alan Stevens 25 Bill Moore 14 Chris Doe 24 Karen Carter 14 Sophie Davis 23 Casey Baker 14 Washington Mutual 23 Art Nelson 14 Fleet Mortgage Co. 21 Cathy Parker 13 31

32 Cluster Visualization 32

33 A Comprehensive Approach Presentation Title

34 A Layered Solution with Fresh Insights

35 Bringing it All Together Medical Claims Analytics Processing Claims Data History Identity Match/ Claims Watch External Evaluators Intelligent Data Retrieval Policy Data Business Rules Predictive Modeling Provider Billing Severity Analysis History Medical Bill Detail PAY Provider Bill Prepayment ClaimFocus SM Evaluation Yes Evaluate Contributory Data IDV & Authentication No Appropriate Claims Handling Process Payer Watch List

36 Thank You! Clint Fuhrman National Dir, Government Healthcare LexisNexis Risk, Inc Linked In Group: LexisNexis Health Care Solutions Twitter: LexisHealthCare Blog: 36