Content uploaded by Gonzalo A Ruz
Author content
All content in this area was uploaded by Gonzalo A Ruz
Content may be subject to copyright.
A Medical Claim Fraud/Abuse Detection System
based on Data Mining: A Case Study in Chile
Pedro A. Ortega
Department of Computer Science
University of Chile
Santiago, Chile
E-mail: peortega@dcc.uchile.cl
Cristi´
an J. Figueroa1
Department of Business Intelligence
Sonda S.A.
Santiago, Chile
E-mail: cristian.figueroa@sonda.com
Gonzalo A. Ruz
Manufacturing Engineering Centre
Cardiff University
Cardiff CF24 3AA, UK
E-mail: ruzg2@cf.ac.uk
Abstract—This paper describes an effective medical claim
fraud/abuse detection system based on data mining used by a
Chilean private health insurance company. Fraud and abuse
in medical claims have become a major concern within health
insurance companies in Chile the last years due to the increasing
losses in revenues. Processing medical claims is an exhausting
manual task carried out by a few medical experts who have the
responsibility of approving, modifying or rejecting the subsidies
requested within a limited period from their reception. The
proposed detection system uses one committee of multilayer
perceptron neural networks (MLP) for each one of the entities
involved in the fraud/abuse problem: medical claims, affiliates,
medical professionals and employers. Results of the fraud
detection system show a detection rate of approximately 75
fraudulent and abusive cases per month, making the detection
6.6 months earlier than without the system. The application
of data mining to a real industrial problem through the
implementation of an automatic fraud detection system changed
the original non-standard medical claims checking process to a
standardized process helping to fight against new, unusual and
known fraudulent/abusive behaviors.
I. INTRODUCTION
Health care fraud and abuse losses represent tens of
billions of dollars each year in many countries [1] [2] [3] [4]
[5]. Medical fraud can occur at various levels [4]. According
to [6], fraud and corruption in health care industry can be
grouped in illicit activities associated to affiliates, medical
professionals, staff and manager, and suppliers. Although
fraud may not necessarily lead to direct legal consequences,
it can become a critical problem for the business if it is
very prevalent and if the prevention procedures are not fail-
safe [7]. There is a difference between fraud prevention and
detection [4]. Fraud prevention describes measures to avoid
fraud to occur in the first place. In contrast, fraud detection
involves identifying fraud as quickly as possible once it has
been committed.
According to the National Health Care Anti-Fraud Associ-
ation, health care fraud is an intentional deception or misrep-
resentation made by a person, or an entity that could result in
some unauthorized benefit to him or his accomplices. Health
care abuse is produced when either the provider practices are
inconsistent with sound fiscal, business or medical practices,
1Corresponding author. Phone: +56-2-6576247; Fax: +56-2-6575160;
Address: Teatinos 500, Santiago-Chile.
and result in an unnecessary cost or in reimbursement of
services that are not medically necessary or that fail to
meet professionally recognized standards for health care [8].
In order to assure the healthy operation of a health care
insurance system, fraud and abuse detection mechanisms
are imperative, but highly specialized domain knowledge is
required. Furthermore, well-designed detection policies, able
to adapt to new trends acting simultaneously as prevention
measures, have to be considered. Data mining which is
part of an iterative process called knowledge discovery in
databases (KDD) [9] [10] can assist to extract this knowledge
automatically. It has allowed better direction and use of
health care fraud detection and investigative resources by
recognizing and quantifying the underlying attributes of
fraudulent claims, fraudulent providers, and fraudulent bene-
ficiaries [11]. Automatic fraud detection helps to reduce the
manual parts of a fraud screening/checking process becoming
one of the most established industry/government data mining
applications [7].
In several countries fraudulent and abusive behavior in
health insurance is a major problem. Fraud in medical
insurance covers a wide range of activities in terms of
cost and sophistication [12]. Health insurance systems are
either sponsored by governments or managed by the private
sector, to share the health care costs in those countries [13]
[14]. Chile has been a pioneer in Latin America initiating
programs which decentralized primary health care services
and encouraged the development of a private health insurance
market in the 1980s [15] [16]. Chilean insured workers and
their dependents may channel their mandatory 7% health
care payroll contributions to either the publicly managed
National Health Fund (FONASA) or to one of the private
pre-paid health insurance plans called ISAPREs (Institutos
de Salud Previsional) [17]. The ISAPREs system, modeled
closely on US HMOs (Health Maintenance Organization),
was created in 1981 [18] [19] with the objective of giving
the consumer more choices through a competitive health in-
surance system and by expanding private provision of health
services. Affiliates of private health insurance companies
may pay an additional contribution for a specific health
plan. The fee determines the level of coverage and a limit
over the health care expenses are defined for each plan. In
1990, the government established a public institution called
Superintendency of ISAPREs [20] [21] to regulate the private
health insurance market. It mediates between consumers and
the ISAPREs, regulates the market in order to guarantee all
contracts, and provides information to the public to increase
the members’ knowledge and market transparency [17].
However, fraud and abuse in medical claims have be-
come an important concern within public and private health
insurance companies in Chile the last years due to the
increasing losses in revenues. In addition, these losses entail
an increase in operation costs diminishing the resources
for health benefits. Indeed, a growth of the subsidies by
labor incapacity and maternity leaves have been registered in
Chile. This increase does not match with an epidemiologist
change of the population, but with an inadequate use of the
health care system [22]. In 1986, a public institution called
COMPIN (Comisi´
on de Medicina Preventiva e Invalidez)
was created by the government for handling complaints of
the affiliates against the ISAPREs [21] [23]. COMPIN has
the faculty for modifying the ISAPRE’s resolution. Doctor-
affiliate complicities are difficult to fight. The labor instability
generates potential unemployment which motivates an illegal
use of these benefits. Objective parameters for the granting
of medical claims do not exist. The exponential increase of
the number of medical claims complicates an appropriate
control. Processing medical claims is an exhausting manual
task that involves analysis, checking, resolution and audit of
high volumes of medical claims daily within a limited period
of three days from their reception. These control activities
are done by few medical experts who have the responsibility
of approving, modifying or rejecting the subsidies solicited.
Therefore, discovering fraudulent and abusive medical claims
is not an easy process for humans experts.
Within the Chilean private health insurance market, Ban-
medica S.A. ISAPRE is the largest company with a 24.2%
market share, delivering health care products, services and
information to 609,514 beneficiaries [20] who have labor re-
lationships with approximately 52,500 employers. Moreover,
Banmedica S.A. has approximately 18,000 medical contracts
with providers and payer organizations. In Banmedica S.A.
the process for checking medical claims was totally non-
standard. For instance, a particular medical claim could be
audited in different ways by two experts. One way is that an
expert reviews the medical claim using non-constant criteria
through time, making it difficult to establish any specific
fraudulent pattern. The other way is based on simple histori-
cal rules elaborated by themselves using limited knowledge.
Therefore, an effective fraud/abuse detection system based on
data mining to detect automatically fraudulent and abusive
medical claims was proposed and implemented in Banmedica
S.A., discovering non-trivial and novelty knowledge and mo-
tivating a pro-active anti-fraud culture within the organization
to fight against new, unusual and known fraudulent/abusive
behavior
In this paper we describe the before mentioned fraud/abuse
detection system that utilizes one committee of multilayer
perceptron neural networks (MLP) for each one of the
entities involved in the fraud/abuse problem: medical claims,
affiliates, medical professionals and employers. This divide-
and-conquer strategy allows to feedback information over
time, combining affiliates’, doctors’ and employers’ behavior.
The paper is organized as follows. Section II delivers
a detailed revision of the literature in relation with the
data mining-based health care fraud/abuse detection. Section
III presents the features of the proposed automatic fraud
detection system. Section IV shows some results about the
evaluation of the proposed employer sub-model. Finally
section V describes the conclusions of this work with some
discussions.
II. RELATED WORK
Phua et. al. [7] highlights fraud committed in insurance
industry as one of the most studied in terms of the number
of data mining-based fraud detection publications, existing
four sub-groups of insurance fraud detection: home, crop,
automobile and medical insurances. In [24] an on-line dis-
counting learning algorithm to indicate whether a case has a
high possibility of being a statistical outlier in data mining
applications such as fraud detection is used for identifying
meaningful rare cases in health insurance pathology data
from Australia’s Health Insurance Commission (HIC). The
performance of a k-Nearest Neighbor (kNN) algorithm with
the distance metric being optimized using a genetic algorithm
was applied in a real world fraud detection problems faced
by the HIC [25]. A multi-layer perceptron (MLP) network
was trained to classify the practice profiles of samples of
medical general practitioners who had been classified by
expert consultants into four classes ranging from normal
to abnormal profiles at the HIC [26]. A set of behavioral
rules based on heuristics and machine learning are used
for performing and scanning a large population of health
insurance claims in search of likely fraud [27]. The hot spots
methodology that entails the use of clustering and rule in-
duction techniques has been used to identify possible frauds
in the Australian Governments public health care system,
Medicare [28]. Becker et. al. identify the effects of fraud
control expenditures and hospital and patient characteristics
on upcoding, treatment intensity and health outcomes in the
Medicare and Medicaid programs [29].
Cox [30] applied a fraud detection system based on fuzzy
logic for analyzing health care provider claims. This fuzzy
system uses rules derived from human experts for detecting
anomalous behavior patterns. Tasks performed in support of
a data mining project for HCFA (Health Care Financing
Administration) such as customer discussions, data extraction
and cleaning, transformation of the database, and auditing of
the data was described in [11] [31]. A data mining framework
that uses the concept of clinical pathways (or integrated
care pathways) was utilized for detecting unknown fraud
and abusive cases in a real-world data set gathered from
the National Health Insurance (NHI) program in Taiwan
[13]. Another model that uses attributes mainly derived from
various expense fields of claims by experts’ consultants was
also designed to detect suspicious claims in the Taiwan NHI
program [32]. An overview of the merging neural networks,
fuzzy logic and genetic algorithms applied to the insurance
industry was done in [33]. Neural networks to classify fraud-
ulent and non-fraudulent claims for automobile bodily injury
in healthcare insurance claims was implemented in [34]. A
method based on naive Bayes that effectively combines the
advantages of boosting and the explanatory power of the
weight of evidence scoring framework was presented in [35].
The method consists of closed personal injury protection
(PIP) automobile insurance claims from accidents that oc-
curred in Massachusetts during 1993 and were previously
investigated for suspicion of fraud by domain experts.
A temporal pattern mining algorithms to identify a set
of frequent temporal patterns gathering insurance claim in-
stances about Pelvic Inflammatory Disease from regional
hospitals in Taiwan was presented [36] [37]. Furthermore,
the classification algorithm C4.5 was applied for fraud/abuse
detection by using the discovered temporal patterns as pre-
dictive features. Some data mining-based approaches which
can be used to extract medical knowledge for diagnosis,
screening, prognosis, monitoring, therapy support or overall
patient management were presented in [38]. Fraud indicators
and rules from the knowledge and experience of human
experts to develop a computer-based expert to facilitate the
work of insurance carriers was carried out by [39]. In Chile, a
single neural network to detect fraudulent medical claims was
implemented in another healthcare insurance company [40].
This scheme utilizes all the data available in arriving medical
claim for constructing a unique vector which is evaluated by
the single neural network.
Several actions that employers can take to reduce losses
from health care fraud, suggesting policy statements and
administrative procedures and guidelines to discourage em-
ployee fraud, to combat provider fraud, to improve health
care fraud detection through the claims payment system and
the possibility of civil and criminal remedies and reviews the
legal theories were pointed out in [12]. Trends in published
neural network application research and exploration of po-
tential neural network research areas were mentioned in [41].
A short discussion of risk and fraud in the insurance industry
was given in [42].
III. METHODS
A. Entities and Medical Claim Data
A medical claim involves the participation of an affiliate,
who is finally the direct benefitted; a medical professional,
who has the faculty for fulfilling the requirement of the
resting period and subsidies for the affiliate; and an employer,
who represents legally the company where the affiliate works.
A private health insurance ISAPRE company like Banmed-
ica S.A. processes daily approximately 800 medical claims.
Each claim is submitted by an affiliate under the approval of
a medical professional justifying the work incapacity. Data
such as age, sex, type of claim, affiliate’s name and date of
birth, ID number, resting period solicited, type and place
of the resting, identification of the medical professional,
identification of the employer, labor activity of the company
where the affiliate works, affiliate’s profession and income
records that the affiliate has gotten in the last three months
are incorporated in each form.
On the other hand, data such as main and secondary
diagnoses, supporting clinical records and medical exams
are also fulfilled by the medical professional, but this data
is sealed when the form arrives to some of the ISAPRE’s
agencies. Once ISAPRE’s medical experts check the forms
they can unveil this data. Just in that moment the medical
experts decide on whether to approve, modify or reject the
number of resting days solicited. In this sense, a neural
classifier that makes a predictive detection of the fraudulent
and abusive claims would be of great help for the medical
experts in their reviewing process, acting as a pre-screen
filter. This predictive detection must only consider historic
data associated to the affiliate, the medical professional and
the employer, and data available before the medical revision
of the arriving medical claim.
B. Business and Data Understanding
Several meetings were held with some medical experts
who explained the main aspects of their criteria for approval,
modification and rejection of medical claims. This allowed us
to better understand the underlying business model, including
discriminative behavioral patterns, as well as weaknesses
of the current non-standarized fraud detection procedure.
The outcome of these meetings was a preliminary set of
variables designed to discriminate between normal and sus-
picious/fraudulent behavior.
Our analysis started with two sets of medical claims.
The first set contained 169 historic abusive medical claims
that were richly documented by Banmedica S.A. until 2003.
These cases manifested notorious fraudulent patterns along
extended periods, hence their manual detection was a simple
task for the medical experts. The second set was a sample of
500,000 labelled medical claims recorded between 2001 and
2003. These were either “approved”, “rejected” or “reduced”.
In order to build an appropriate set for the application of
machine learning techniques, both sets had to be completed
by computing their remaining descriptive features from the
ISAPRE’s relational database. In particular, a subset of nine
tables was used. These tables had data associated to each
entity involved in a medical claim as for instance payment
behavior, medical resolutions, frequencies, beneficiaries and
historic incomes. Consequently, a single data set was cal-
culated and stored into a unique repository for further data
analysis.
Next, several tasks to assess the data quality were done. We
checked for correlations among features and discriminative
power, as well as their consistency with empirical knowledge.
For instance, seasonality of medical claim submission with
respiratory, psychiatric, maternal leaves diagnoses, correla-
tions between days submitted and effective days of subsidy
and income and payroll contribution analysis were carried
out. Finally, a total amount of 125 features were selected.
However, this analysis also showed that the majority of
the 169 fraudulent/abusive documented medical claims were
concentrated around a reduced number of medical profes-
sionals, affiliates and employers. For example, 19 employers
and 6 doctors were implicated with 152 medical claims. Fur-
thermore, due to the lack of standards and the huge amount
of medical claims, the labelling of the second data set was not
sufficiently accurate, and in some cases even contradictory.
Moreover, 35% of these medical claims belonging to year
2001 were discarded due to their poor quality in terms of
missing values and low contribution. Also, medical claims
whose number of solicited resting days was lower than 10
were excluded, according to their almost negligible impact
in profits.
The poor quality of the data forced us to rebuild the data
set by relabelling a subset of cases and to apply aggressive
dimensionality reduction techniques.
C. Data Preparation
In order to build a robust classifier using only a small
training set, it was decided to apply a divide-and-conquer
strategy. The initial problem was subdivided into smaller
problems, namely, four separate models to cope with the
entities before mentioned. Each sub-model required a smaller
number of features and training samples. The detailed de-
scription follows.
First, an exhaustive manual classification was done, as-
sisted by both medical experts and legal advisors. For each
entity, that is, medical claims, affiliates, medical profession-
als and employers, training sets of sizes 2838, 424, 590 and
394 samples were obtained. Each set had an equal proportion
of fraudulent/abusive and normal cases.
Second, the feature vectors for the four different models
were further optimized. A manual feature selection procedure
was applied to reduce the feature vectors. For example,
for groups of highly correlated features, only the most
discriminative one was chosen. Furthermore, a categorical
feature was avoided or replaced by a continuous feature
whenever possible. Intimately related but complementary
variables were fused into non-human interpretable features
that are better suited for a continuous classifier. The overall
maxim has been to design robust features that summarize
temporal behavior over an extended time span. Most of them
involved historical data, generally moving windows of 12
months backwards starting from the submission date. As a
result, the medical claim, affiliate, medical professional and
employer sub-models had feature vectors of sizes 14, 25,
17 and 12 respectively. As an example, Fig. 1 shows the
histogram of the variable expected resting days approved by
the employer for the 394 manually classified employer cases,
separated by fraudulent/abusive (F) and normal (N) cases.
Finally, standard data preparation techniques have been
applied to avoid training biases. These include the removal of
2% of the outliers and a linear normalization of the features.
The data transformation parameters have been saved for their
later use during production stage.
0 50 100 150 200 250 300 350 400
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Value
Training Sample
N F
Fig. 1. Histogram of the variable expected resting days approved by
the employer for the 394 manually classified employer cases, separated by
fraudulent/abusive (F) and normal (N) cases
D. Modelling
The proposed system allows the identification of fraudu-
lent and abusive behavior in medical claims, affiliates, med-
ical professionals and employers. Upon the medical claim
submission, the different entities are analyzed separately
using historical data with cross-references among them. As
mentioned above, the predictive model cannot make use of
the sealed diagnostic information contained in the medical
claim for its decision, although historical data is incorporated
via the features.
TABLE I
EMPLOYER MODEL, CLASSIFICATION VARIANCE
Classifier Mean Std. Dev.
Single MLP 88.7 % 8.4 %
Committee (10 MLPs) 88.7 % 1.8 %
Initially, each sub-problem was modelled by a standard
two-layer neural network (MLP). Albeit architectures were
kept simple (e.g. using small hidden layers), the classifiers’
accuracy showed a surprisingly high variance that exceeded
8% between different runs. In order to decrease the model
variance, we decided to implement a committee of mul-
tilayer feedforward neural networks [43][44]. Alternative
complexity reduction techniques such as network pruning
were discarded due to the high weight variance even for very
simple architectures. In the case of committees, the output
of several independently trained networks was averaged in
order to reduce their model variance, thus delivering robust
results specially when the number of samples is low. Each
sub-model had committees of 10 MLPs. Table I compares
the standard deviation obtained for single network against
a committee. The chosen architectures were 14-3-1, 25-3-1,
17-3-1 and 12-3-1 for the medical claim, affiliate, medical
professional and employer committees respectively. All ac-
tivation functions were sigmoidal. The output unit indicates
whether the sample corresponds fraudulent/abusive or normal
behavior. For each case, the labelled data set were divided
into a training set, a validation set and a testing set. The
training procedure minimized the error measured on the val-
idation set while using the training set to adjust the networks’
weights. This technique known as early stopping provides a
principled method for selecting models that generalize well
without sacrificing capacity, hence avoiding over-fitting to
the training data’s noise while keeping the classifier’s ability
to learn non-linear discriminant boundaries [45]. The testing
sets were used to estimate the generalization performance of
the system. It is worth noticing that the design strives for
model simplicity and improved generalization ability under
a limited sample set scenario.
Our fraud detection system is illustrated in Fig. 2. This
diagram shows the four sub-models, namely, a committee
of neural networks for each entity. In each case, the inputs
are precomputed feature vectors that model the particular
fraud and abuse subproblem. The output of each committee
delivers a predictive value each time that a medical claim is
received by the ISAPRE. These values serve as additional
inputs to the sub-models, providing a feedback mechanism
for combining the different results. For instance, the outputs
of the affiliate, medical professionals and employer models
provide strong and condensed evidence for the medical claim
sub-model. Models are evaluated at fixed time intervals
according to a predefined schedule. The medical claim model
is executed daily to process incoming submissions, whereas
the other are executed once monthly.
Fig. 2. Fraud detection scheme by using sub-models with feedback
connections.
All sub-models are retrained monthly. In order to keep
the training samples representative for historical and novel
fraud behavior, a data renewal procedure has been defined.
New training samples are chosen randomly and rigorously
classified by experts. A subset of equal proportions of fraud
and normal cases are then selected and incorporated to
the training set. This way, the model is kept updated and
informed of novel fraud trends, acting simultaneously as a
prevention mechanism.
E. Incorporation into Fraud Detection Workflow
Approximately 800 digitalized medical claim forms arrive
during each day. Our automated fraud detection system is ex-
ecuted at night, assigning a fraud probabilities to each form.
Associated affiliates, medical professionals and employers
are updated as well. At the next day, a web interface allows
to consult these records ranked by their fraud probabilities,
acting like a pre-screen filter.
Several costs are involved in this fraud detection pro-
cedure, such as the personnel’s salaries, false alarms and
fraud or abuse cases that are not detected. Based on this
information, an optimum fraud probability threshold can
be estimated. Records with lower fraud probabilities can
be excluded from this revision procedure. The description
follows.
As mentioned in [46], a naive classifier can get high
accuracy on a skewed data set, where the prevalence of
fraud is very low, but at the cost of misclassifying all the
true fraudulent/abusive cases. Therefore, an optimum deci-
sion threshold was estimated by plotting receiver operating
characteristic (ROC) curves [47] instead of the classifier
accuracy. ROC curves plot the true positive (TP) against
the false positive (FP) rates at different decision making
thresholds and determine the test’s ability to differentiate
between fraudulent/abusive and normal cases. TP and FP
are the per unit rate of correct and false fraudulent/abusive
classifications respectively. Moreover, an iso-performance
line was plotted in the TP −FP plane. Classifiers on this line
all have the same expected cost [48]. Two points (TP1, FP1)
and (TP2, FP2)have the same performance if
TP2−TP1
FP2−FP1
=p(N)C(FP )
p(P)C(FN),(1)
where p(N)and p(P)are the prior probabilities of obtaining
a negative and a positive example respectively, and FN is the
false negative rate. The C(FP )and C(FN)represent the
costs of a FP and a FN error, respectively. This equation
defines the slope of the iso-performance line. Given a ROC
curve and an iso-performance line, the point of intersection
defines the optimal operation point.
IV. RESULTS
The results associated with our fraud detection system
such as classifier accuracy and total savings are confidential
due to a disclosure contract. Nevertheless, Banmedica S.A.
has granted the permission to publish results related to the
employer sub-model.
Historical data considering a total of 8,819 employers was
analyzed. This set contains 418 fraudulent/abusive and 8,401
normal cases. This data set was divided into a training,
validation and test set (Table II). For training purposes, the
fraud/abuse cases in both the validation and training sets were
quadruplicated to remove possible bias.
Figure 3 shows the ROC curve obtained with the prediction
employer sub-model. The top left point (0,1) represents
TABLE II
EMPLOYER DATA SET
Employer Category Number of Cases Percentage
Fraud/abuse (T) 176 20.0
Normal (T) 706 80.0
Total (T) 882 100.0
Fraud/abuse (V) 118 20.1
Normal (V) 470 79.9
Total (V) 588 100.0
Fraud/abuse (T) 124 1.7
Normal (T) 7,225 98.3
Total (T) 7,349 100.0
the perfect predictor where all fraudulent/ abusive cases
are caught without generating false alarms. Therefore, the
closer the ROC curve is to the point (0,1) the better is
the performance. To plot the iso-performance line, the prior
probabilities have been estimated using random sampling
and manual inspection (based on data between June 2005
and January 2006), obtaining p(P) = 0.017 and p(N) =
0.983. Furthermore, a scenario where the cost of a FN error
corresponds to at least 5 times the cost of a FP error has been
setup, yielding a cost ratio of C(FP )/C(FN )>0.2. At the
operation point the TP and FP rates were 73.4% and 6.9%,
respectively, i.e., the prediction module was able to identify
73.4% of the true fraudsters/abuses, screening only 8.7% of
all employers. Although there were 6.9% false positives, this
is acceptable for a model which is designed to work as a
pre-screen filter. Those identified as fraudsters/abuses by the
committee are then referred for a detailed revision.
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
False Positives (FP) rate
True Positives (TP) rate
Iso−performance line
ROC curve
Fig. 3. ROC curve
To illustrate the efficacy of our fraud detection system,
Table III shows the evolution of an employer’s fraud score.
In this case, an employer whose first record in the IS-
APRE is saved on September 13, 2005 is analyzed. The
employer ID number is 7771xxxx. This first record originated
from an associated affiliate’s medical claim submission. The
predictive employer sub-model was only able to deliver a
value 0.61 of being fraudulent or abusive because there is
no history associated to the employer until that point. The
second medical claim arrived to the ISAPRE on October
13, 2005. After the update, employer sub-model scored this
record with an abuse probability of 0.69. Although this score
increased since the first submission, it did not reach the
decision threshold. After a third submission on October 18
and a fourth on October 31, the score augmented to 0.82 and
0.91 respectively, triggering its manual revision procedure.
The experts confirmed the abuse case and activated an alarm
rule that highlights future medical claims from this employer.
Consequently, our fraud detection system was able to identify
this case within only 2 months. This single example is
representative: based on historical data, our fraud detection
scheme identifies fraud/abuse employer behavior within 2
months, whereas the former procedure took 8.6 months in
the average.
TABLE III
FRAUD SCORE EVOLUTION,EMPLOYER ID 7771XXXX
Fraud/abuse probability Reception date of medical claim
0.6118 September 13, 2005
0.6908 October 13, 2005
0.8246 October 18, 2005
0.9126 October 31, 2005
V. CONCLUSIONS AND DISCUSSION
With the implementation of our fraud detection system in
medical claims, a proactive anti-fraud culture was generated
within Banmedica S.A. to fight against new, unusual and
known bad behaviors filtering opportunely suspicious med-
ical claims. Insightful knowledge has been gathered, allowing
to build a fraud/abuse taxonomy that identifies 15 different
types. Moreover, this automatized revision procedure moti-
vated improvements in the manual revision process.
During 2005 Banmedica S.A. was able to estimate that
the new fraud detection system rejected medical claims that
contributed between 9.5% and 10.0% of the overall raw costs.
Additionally, persistent fraud and abuse cases are detected
6.6 months earlier than without the system, as estimated from
historical records. The savings are considerably higher due
to the premature detection of these cases
Banmedica S.A. has changed its fraud detection policy
from its former reactive detection strategy, which yielded
only a few documented cases up to 2003, to a proactive
and preventive detection strategy, reaching approximately 75
fraudulent and abusive cases each month during 2004 and
2005. This drastic increase is mainly due to the contribution
of our fraud detection system. The current fraud detection
rate yields savings that cover operational costs and allowed
to increase the quality of the health care coverage, fully
justifying the investment.
Unlike other fraud detection systems that are based on a
single monolithic model, our divide-and-conquer approach is
able to analyze each involved entity separately, along with
other beneficial side-effects such as dimensional reduction
and the consequent model robustness. It is worth noticing
that the detection of the sources of fraudulent and abusive
behavior is a far more efficient strategy than the analysis
of the individual medical claims. The detection of a source,
namely, an affiliate, a medical professional or an employer,
allows to cover multiple present and future submissions.
Future work considers the design of an improved model
to estimate costs and savings associated to the operation of
our detection scheme, as well as to measure its preventive
secondary effects.
ACKNOWLEDGMENT
The authors would like to thank Arturo Phillips, Jaime
Ochagav´
ıa and Ver´
onica Rodr´
ıguez who belong to the health
management of Banmedica S.A. for their valuable help and
support during the implementation of the model. We also
thank Johanna Guzm´
an, V´
ıctor Marchant, Cristi´
an R´
ıos, Juan
Pablo Herrera, Mauricio Marcos and Andr´
es Vergara who
belong to the Business Intelligence Department of Sonda
S.A. for their help in building the repository and managing
different activities that allowed the proper installation of the
fraud detection system.
REFERENCES
[1] United States General Accounting Office, “Health insurance: Vulnerable
payers lose billions to fraud and abuse,” GAO-HRD-92-69, 1992.
[2] M. Lassey, W. Lassey and M. Jinks, “Health care systems around the
world: Characteristics, issues, reforms,” Englewood Cliffs, NJ: Prentice-
Hall, 1997.
[3] S. J. Hong and S. M. Weiss, “Advances in predictive models for data
mining,” Pattern Recognition Letters, vol. 22, pp. 55–61, 2001.
[4] R. J. Bolton and D. J. Hand, “Statistical fraud detection: A review,”
Statistical Science, vol. 17, no. 3, pp. 235–249, 2002.
[5] Health care fraud, “Health care fraud: A serious and costly reality for all
Americans,” National Health Care Anti-Fraud Association (NHCAA),
REPORT, http://www.nhcaa.org, April 2005.
[6] European healthcare fraud and corruption network,
http://www.ehfcn.org/, 2005.
[7] C. Phua, V. Lee, K. Smith, and R. Gayler, “A comprehensive survey
of data mining-based fraud detection research”, Artificial Intelligence
Review, submitted, 2005.
[8] Guidelines to health care fraud, “Guidelines to health care fraud,” Na-
tional Health Care Anti-Fraud Association (NHCAA), NHCCA Board
of Governors, http://www.nhcaa.org, 1991.
[9] U. M. Fayyad, “Data mining and knowledge discovery: making sense
out of data,” IEEE Expert, Intelligent Systems and their Applications,
pp. 2025, October 1996.
[10] J. Han and M. Kamber, Data mining: Concepts and techniques, Morgan
Kaufmann Publishers, San Francisco, 2001.
[11] L. Sokol, B. Garcia, J. Rodriguez, M. West and K. Johnson, “Using
data mining to find fraud in HCFA health care claims,” Top Health
Information Management, vol. 22, no. 1, pp. 1-13, 2001.
[12] B. B. Pflaum and J. S. Rivers, “Employer strategies to combat health
care plan fraud,” Benefits Quarterly, vol. 7, no. 1, pp. 6-14, 1991.
[13] W. S. Yang and S. Y. Hwang, “A process-mining framework for
the detection of healtcare fraud and abuse,” Expert Systems with
Applications, Article in Press, corrected proof, 2005.
[14] L. J. Opit, “The cost of health care and health insurance in Australia:
Some problems associated with the fee-for-service,” Soc. Sci. Med.
vol. 18, no. 11, 967–972, 1984.
[15] J. Jimenez and T. Bossert, “Chile’s health sector reform: lessons from
four reform periods,” Health Policy, vol. 32, pp. 155–166, 1995.
[16] A. Viveros-Long, “Changes in health financing: The Chilean experi-
ence,” Soc. Sci. Med. vol. 22, no. 3, pp. 379–385, 1986.
[17] F. Bertranou, “Are market-oriented health insurance reforms possible
in Latin America?. The cases of Argentina, Chile and Colombia,” Health
Policy, vol. 47, pp. 19–36, 1999.
[18] E. Miranda, J. Scarpaci and I. Irarr´
azaval, “A decade of HMOs in
Chile: market behavior, consumer choice and the state,” Health & Place,
vol. 1, pp. 51–59, 1995.
[19] C. Sapelli and B. Vial, “Self-selection and moral hazard in Chilean
health insurance,” Journal of Health Economics, vol. 22, pp. 459–476,
2003.
[20] Superintendencia de ISAPRES, October, 2005
http://www.superintendenciadesalud.cl/.
[21] Ley N 18.933 Publicada en el Diario Oficial de 09.03.90 Crea
Superintendencia de ISAPRE, dicta normas para el otorgamiento de
prestaciones por ISAPRE y deroga el D.F.L. N 3, de 1981, del
Ministerio de Salud, Chile.
[22] P. Melero, “Cara Evidente de Licencias M´
edicas: Incentivos Inver-
tidos,” Comisi´
on de Salud, C´
amara de Diputados de Chile, Agosto,
2005, available at http://www.isapre.cl.
[23] Decreto supremo N 42, 1986, Ministerio de Salud, Chile.
[24] K. Yamanishi, J. Takeuchi, G. Williams, and P. Milne, “On-line
unsupervised outlier detection using finite mixtures with discounting
learning algorithms,” In Data Mining and Knowledge Discovery vol. 8,
pp. 275–300, 2004.
[25] H. He, W. Graco, and X. Yao, “Application of genetic algorithms and
k-nearest neighbour method in medical fraud detection,” In Proc. of
SEAL1998, pp. 74–81, 1999.
[26] H. He, J. Wang, W. Graco and S. Hawkins, “Application of neural net-
works to detection of medical fraud,” Expert Systems with Applications,
vol. 13, no. 4, pp. 329–336, 1997.
[27] J. Major, and D. Riedinger, “EFD: A hybrid knowledge/statistical-
based system for the detection of fraud,” Journal of Risk and Insurance
69(3), pp. 309-324, 2002.
[28] G. Williams, “Evolutionary hot spots data mining: An architecture for
exploring for interesting discoveries,” In Proc. of PAKDD99, 1999.
[29] D. Becker, D. Kessler and M. McClellan, “Detecting medicare abuse,”
Journal of Health Economics, vol. 24, pp. 189–210, 2002.
[30] E. Cox, “A fuzzy system for detecting anomalous behaviors in
healthcare provider claims,” In Goonatilake, S. & Treleaven, P. (eds.)
Intelligent Systems for Finance and Business, pp. 111-134. John Wiley
and Sons Ltd., 1995.
[31] L. Sokol, B. Garcia, M. West, J. Rodriguez and K. Johnson, “Pre-
cursory steps to mining HCFA health care claims,” Proceedings of the
Hawaii International Conference on System Sciences, Hawaii, pp. 6019,
2001.
[32] C. L. Chan and C. H. Lan, “A data mining technique combining fuzzy
sets theory and Bayesian classifier – An application of auditing the
health insurance fee,” Proceedings of the International Conference on
Artificial Intelligence, pp. 402–408, 2001.
[33] A. F. Shapiro, “The merging of neural networks, fuzzy logic, and
genetic algorithms,” Insurance: Mathematics and Economics, vol. 31,
pp. 115–131, 2002.
[34] P. L. Brockett, X. Xia, and R. A. Derrig, “Using Kohonen’s self-
organizing feature map to uncover automobile bodily injury claims
fraud,” The Journal of Risk and Insurance, vol. 65, no. 2, pp. 245–
274, 1998.
[35] S. Viaene, A. Richard and D. G. Dedene, “A case study of applying
boosting Naive Bayes to claim fraud diagnosis,” IEEE Transactions on
Knowledge and Data Engineering, vol. 16, no. 5, pp. 612–620, May
2004.
[36] C. P. Wei, S. Y. Hwang and W. S. Yang, “Mining frequent temporal
patterns in process databases,” Proceedings of International Workshop
on Information Technologies and Systems, Australia: Brisbane, pp. 175-
180, 2000.
[37] S. Y. Hwang, C. P. Wei, and W. S. Yang, “Process mining: Discovery
of temporal patterns from process instances,” Computers in Industry,
vol. 53 no. 3 pp. 345-364, 2004.
[38] N. Lavrac and N. Lavrac, “Selected techniques for data mining in
medicine,” Artificial Intelligence in Medicine, vol. 16 no. 1, pp. 3-23,
1999.
[39] W. Herb and M. Tom, “A scientific approach for detecting fraud,”
Best’s Review, vol. 95, no. 4, pp. 78–81, 1995.
[40] C. Cooper, “Turning information into action,” Computer Associates:
The Software That Manages eBusiness, Report, January 2003, available
at http://www.ca.com.
[41] B. K. Wong, T. A. Bodnovich and Y. Selvi, “Neural network appli-
cations in business: A review and analysis of the literature (1988-95),”
Decision Support Systems, vol. 19, pp. 301–320, 1997.
[42] B. Glasgow, “Risk and fraud in the insurance industry,” AAAI Work-
shop: AI Approaches to Fraud Detection and Risk Management, AAAI
Press, Menlo Park, California, pp. 20–21, 1997.
[43] M. P. Perrone, “General averaging results for convex optimization”,
Proceedings 1993 Connectionist Models Summer School, pp. 364-371,
Hillsdale, NJ: Lawrence Erlbaum, 1994.
[44] M. P. Perrone, L. N. Cooper, “When networks disagree: ensemble
methods for hybrid neural networks”, Artificial Neural Networks for
Speech and Vision, pp. 126-142, London: Chapman & Hall, 1993.
[45] R. Caruana, S. Lawrence, C. L. Giles, “Overfitting in neural networks:
backpropagation, conjugate gradient, and early stopping”, Neural Infor-
mation Processing Systems, Denver, Colorado, 2000.
[46] P. A. Est´
evez, C. M. Held and C. A. Perez, “Subscription fraud pre-
vention in telecommunications using fuzzy rules and neural networks,”
Expert Systems with Applications, vol. 31, no. 2, 2006, (in press).
[47] T. Fawcett, “ROC graphs: Notes and practical considerations for data
mining researchers,” HP Labs. Technical Report HPL-2003-4, 2003.
[48] F. Provost and T. Fawcett, “Analysis and visualization of classifier per-
formance with nonuniform class and cost distributions,” AI Approaches
to Fraud Detection & Risk Management, Workshop Technical Report
WS-97-07, AAAI Press, Menlo Park, California, 1997.
