Transcript Detecting Suspicious Claims

Detecting Suspicious
Claims :
An Operational
Perspective
Marty Ellingsworth
Director, Operations Research
Customer Research and Strategies
Fireman’s Fund Insurance Company
November 14, 2001
National Insurance Crime Bureau (NICB)
Most common insurance fraud scams of 2000.
Each has its own set of features for detection - if you can find them.
Bodily Injury Fraud
•
often associated with staged or caused auto accidents
•
involve fabricating physical injuries
•
often with dishonest doctors and lawyers (conspiracy and collusion)
Auto Repair Fraud
•
claimant gets high appraisal, in cooperation with an unscrupulous repair sh
•
gets a vehicle repaired; pocketing the difference
Homeowners Claim Fraud
•
arson for profit
•
fabricating claims for phony burglaries
•
padding of legitimate claims for theft or damage to the home
Workers Compensation Fraud
•
faking injuries or exaggerating the extent of a minor injury
•
claiming work relatedness for an injury sustained at home
Exercise on problem solving: Move a tree
• Define the problem
• Formulate a solution
• Get the ‘right’ tools
• Learn how to use them
• Adapt to the situation
• Assess the results of your actions
• Make improvements
Taking action on an individual claim can be challenging.
Fraud and Abuse: Problem Segmentation
Many current data tracking systems can not delineate the specific behavior(s) that resulted in
the claim going to the SIU, nor its ultimate outcome. Because of this, all of these different
‘signals’ get lumped together for modeling historical SIU as Yes/No.
Suspicion - No
Evidence
Build-up Found
• Claimant Opportunistic Build-up
• Exaggerating / Padding / Inflating / Rounding
'Hard' Fraud
Found
No Suspicion
• Planned / Staged Accident
• Attorney/Provider Collusion
• False Billing
• Previous Damage/Injury
• Faking Disability
• Not related to the accident
• Fictional Claims
• Premium Fraud
• Adjuster Fraud
• Agent Fraud
• Identity Fraud
• Organized Crime
Look for fraud in the “ Life of a Claim “
Fraud and abuse can occur at any time during a claim.
?
?
??
Presumed Legitimate
Investigate
Inflammatory Red Flag
Being on “Watch List”
False Identity
Stolen goods
Faked the Loss
Caused the Accident
Evaluate
Negotiate Settle
Claimant “Build -up”
Padded Estimates
Exaggerated Lost Earnings
Multiple Red Flags
Collusion, Conspiracy,
Extensive Claims History
False/exaggerated Disability
Match to ‘Bad Guy’ Data base
Large Data base Link Analysis
Connected to a Crime Ring
How do we detect fraud and abuse?
“Adjuster Centric” referral systems often do not collect data electronically and frequently do not
get applied consistently between adjusters over the life of the claim. Many different methods of
intelligent data gathering and analysis can be successfully employed for effectively detecting fraud.
• Training Adjusters
• Claim-based “Red-Flags”
Manual, On-line, or Batch Processing
• Database Submissions and Searches
Automated, Directed
• Expanded Data collection and feedback of claim outcomes
• Expert Systems (Bill Review) and Business Rules
• Statistical Modeling
Likelihood, Outliers, Dissimilarity within latent groups,
Variance from Expected Behavioral ‘Signature’
• Visualization
Timelines, Geographic mapping
• Link Analysis (especially with Industry Databases)
What can we do to resist fraud and abuse?
If a claim is suspected of ‘hard’ fraud, then we should work with Federal, State, and Local authorities
to resolve the claim - both criminal and civil issues. In many cases, build-up claims are negotiated
by the adjuster after considering the medical damages submitted. Oftentimes, the SIU is not involved,
or it is notified after all of the medical treatment has accumulated.
With Evidence of Fraud:
Negotiation: Build-up Cost Reduction
Effective medicals management can assist in reducing build-up, by automatically identifying
and flagging claims with irregular treatment as compared to normative treatment patterns.
Close Claim with no payment
Refer Claim to Authorities
Assist in criminal prosecution
Seek civil damages
Medical management can also be used to identify and flag claims with
potentially fraudulent medical treatment, such as...
•
Ove rpricing:
•
Irregula r Treatment:
•
Unrelated Treatment:
•
Exc ess Utiliz ation:
Service fe es consistently or significantly above
norm ative rates and fees
irregular/ inappropriate me dical treatment patterns, such
as high dia gnostic $ to c ura tive $ ratio or high $ charges
within short LOE
medical services generally not associate d with
underlying accident-related injury
excessively frequent medical services a s com pared to
norm ative benchmarks (80th pe rcentile or 95th
percentile)
If a claim shows any of the above irregularities:
• Investigate claim further to determine if irregularities are
warranted (e.g., claimant is pregnant)
• Request IME
• Use treatment irregularities as leverage point during
negotiation process
Lessons Learned
We need automated referrals made on timely, accumulated information to be most
effective in resisting fraud and abuse, and to get the most efficient productivity from
our resources.
•Data mining can add considerably to the Manual /
communication methods now in place
•Time is of the essence for making an impact on treatment
•Big hurdle in initially building a data set for analysis
• Company skill set, hardware, and dedicated resources
• Some important factors were not historically collected
•Text Mining as an information extraction tool is quite valuable
•Fielding sophisticated models can depend significantly on IT
•Continue to collect feedback on referrals to improve models over time
•Industry data would be useful for moving beyond claimants.
Case Example: Auto 3rd Party BI
Business Objective
Reduce Unnecessary Losses Paid Due to Fraudulent and Abusive Claims
Increase Efficiency of SIU Resources
-
Sharpen our recognition of potentially fraudulent claims (find more claims)
Shorten the time it takes to get an SIU resource involved (find them quicker
Reduce unqualified referrals generated by quotas
Reduce time spent by SIU staff on training adjusters
LIKELY ACTION STEPS
Interdiction of build-up during treatment
Negotiate ‘Build-Up’
Litigate ‘Hard Fraud’
3rd Party ABI Fraud - Classification of Suspicion
• The historical data show suspicion (1) or
not suspicion (0) as indicated by the SIU’s
non-administrative presence in a claim
• Many different methods can be applied to
rank order claims to differentiate highly
suspicious claims from not suspicious.
• We decide to send a claim for SIU review
based on a precision criteria
• “Fast Track” claims can also be filtered
• Precision and Recall criteria can be
balanced to SIU resource availability
Predicted
100%
Fraud
Suspicion
Fraud
No Fraud
True Positives
(want to
maximize)
False
Negatives
(want to
minimize)
False
Positives
(want to
minimize)
True
Negatives
(want to
maximize)
of
Observed
Fraud
No
Fraud
0%
Low
High
Level of Independent Variable(s)
Data Mining Methods
The complex resources needed to attack many of the fraud segments leave many insurers using ‘low tech’
red-flag systems and emphasizing better communication between adjusters and investigators.
Data mining adds summarized data information to the process for better results.
Conditional
Logic
Discovery
Patterns and
Associations
Trends and
Variations
Data Mining
Predictive
Modeling
Forensic
Analysis
Outcome
Prediction
Deviation
Detection
Link Analysis
What makes a Claim look suspicious?
Across business lines similar themes evolve which highlight claimant behavior associated
with fraud and abuse claims. Non-claimant fraudsters are much more difficult to pinpoint.
• INCONSISTENCY
• DENYING PRIOR CLAIM HISTORY
• UNCOOPERATIVE // TOO COOPERATIVE
• TIME LINE OF EVENTS
• DETAILS FOR SETTLEMENT (TOO MANY/TOO FEW)
• CIRCUMSTANCES UNLIKELY
Data Exploration: Secret of the Red Flags
A few of the Red-Flags have individual strength in indicating the need for
an SIU triage, but most are in the 15 - 30 % ‘Hit Rate’ range. Combining
responses into answer vectors can dramatically increase the ‘Hit Rate’.
YYNNYY
= 85%
Claimant is demanding an unusually quick settlement
18%
Claimant is unusually familiar with insurance terms / procedures
12%
Multiple unrelated claimants were represented by the same attorney
35%
The claimant’s vehicle was damaged in a prior accident
13%
The claimant has been involved in other accidents in the past 3 years
21%
The facts of the accident cannot be confirmed
16%
The insured felt set up
19%
Medical bills lack the detail needed to properly evaluate the claim
14%
Claimant refuses to provide information or submit to an IME
16%
Claimed injuries are inconsistent with the facts of the accident
20%
Treatment received is inconsistent with the claimed injuries
32%
What it takes to Create the Data Set
Business Line Exec
Field Office Staff
Claims Trainers
CLAIMS
Data
Analyst
MEDICAL
COST
CONTAINMENT
Data Collection:
I S and Analyst
Project Leader
SPECIAL
UNDERWRITING INVESTIGATION
UNIT
POLICY DATA
RECOVERY
UNIT
Datamart
Programmer
Domain Expert
Knowledge:
- Auto
- GL
- Property
- Work Comp
Personal
Commercial
What to learn from Structured Data
Significant pre-processing of raw data is needed for creating useful informational features out
of existing structured data. Rolling-up payment transactions, and collecting and integrating
detailed medical bill data with the claim data can result in powerful predictive variables.
• Repeatable Patterns
• Trends, Seasons, Cycle
• Propensities, Likelihood
• Causation and Interaction
• Ratios between Dollars and Distances
• Stakeholder Behavior
• Unlikely Occurrences
Sophisticated Transformation of Data
Data mining end-work-product data record is optimized for outcomes analysis.
In this case, everything is rolled up/down to the third party claimant.
Claim / Policy / Development / Review / Treatment / Savings /Fees / Provider
Claim Master File
Policy System File(s)
Claim Payment Detail File
Claim Reserve History File
Bill Review Header File
Supplemental Sources
ISO, NICB,
Litigation Sub-system
Bill Review Bill Detail File
Provider/Vendor File(s)
The Claims “Checkbook”
By integrating our claims data with our medical bill review vendors’ data, we can see to
whom, when, and where our money is going. This ‘follow the money’ process will give us
the details for tracking patterns of collusion in our claims, but with only a fraction of the
market share, we’ll need to access Industry data to identify organized rings.
Claim System
Claim File
$x,xxx.xx
Payments
Medical Payments
Indemnity Payments
Expense Payments
Reserves
Bill Review Vendor
Medical Bill Review Systems
Bill Record
Bill Line Item Detail
Reduction Reasons
Charged versus Paid
• Bill Review Rule
• Fee Schedule
• U&C Repricing
• PPO Discount
• Other Savings
Bill Review Rule Reasons
Use Review Reduction reasons for negotiating damages.
What to learn from Unstructured Data
To segment types of fraud and to baseline which Red Flag questions help the most, you can
process the ‘free text’ fields in the claim administration system. Both “Text Mining” and
Natural Language Processing methods can extract actionable information from text data.
• Claim file coding leaves a lot to be desired.
• Powerful new variables can be created for millions of claims
without the cost and time lag of manual review
• Notes in the file are indicative of events of special interest
- suspicious behavior
- legal representation
- subrogation opportunity
- injury severity
• Notes are “time stamped” so we can see chronologies
Text Mining Task - Extracting Information
from unstructured data in the claim file
DATE
DATE OF
OFLOSS
LOSS 11/07/99
PROGRESS
NOTE
PROGRESS NOTE
“felt set up”
“suddenly stopped”
Repped in less
than 4 Days
ProcID
Name
Date
ProcID
Name
Date
409F123
Ima Phile-Hanler 11/11/99
“Insured said that they felt set up, this was a mild
impact in heavy traffic that happened when the
claimant suddenly stopped while other traffic kept
moving. Claimant is represented.”
DESCRIPTION
LOSS
DESCRIPTIONOF
OF
LOSS
“ Minor RE in Heavy Traffic”
RESERVES
RESERVES ABI $7500
Inconsistency
Minor Impact v.
Severe Injury
Current Detection Capability: Auto 3rd Party BI
Using the best of structured and unstructured data features we are able to create a very strong rank
ordering of cases for the SIU to review. In our research, 76% of all the historical ‘Bad Guys’ for
Auto 3rd Party Bodily Injury claims are found in the top 15% of the ranked cases.
Suspicion Level by Score Group
100%
5%
90%
80%
70%
91%
60%
50%
40%
30%
20%
10%
0%
1
2
3
4
Not Suspicious
5
Suspicious
6
7
8
Where to next? Change the paradigm
A claim can be considered a document where information is added over time until complete.
Breakthrough thinking -- you can dynamically route claims much like a newswire
subscription service classifies and routes in-coming stories, or like an internet search engine
finds web-sites which have the content you want to see.
Field additional rules and scoring engines. Search for more powerful predictors.
Continuous collection of data and feedback of results.
Extend the practical ability of classifying claims using text mining indexing strategies.
Pursue using ‘web spidering’ technology to combine information extraction
enhanced models with real-time indexing of claim notes for fast and efficient recognition
of claims of interest.
Integrate feedback loops for a spider based inference engine to dynamically
route claims based on emerging information in the file
For non-claimant fraud, we will explore methods to combine information with
larger data sets to better enable data mining techniques to reach the next level.
Name and address standardization and parsing is needed, and ‘similarity’ engines
will be invaluable for finding people trying to hide their identities.