Insurance fraud is one of the most expensive economic financial crimes. Most risk management solutions use rules to detect potential abuse, but as the patterns of abuse change, those solutions become inefective. In this paper we apply machine learning (Decision Trees, Bagging, Random Forests and Boosting) for fraud detection in health insurance. Performance of the model is evaluated using accuracy, error rate, sensitivity and specificity. The best results were achieved using Bagging technique. In further research it would be useful to analyze applicability of deep learning models and anomaly detection methods.
Insurance fraud is one of the most expensive economic The health insurance fraud claims are broadly classiifnancial crimes [1]. In order to conquer a larger share ifed into several classes. In [2] a solution for dupliof the insurance market, insurance companies ofer cated claims fraud is proposed, namely, detection of beneficial insurance terms and, as a result, in such a cases, when people are submitting just slightly diferway providing new opportunities for fraud as well. ent bills repeatedly, changing some small portion like The number of crimes in the insurance sector is in- the date, in order to charge insurance company twice creasing every year. As a result, service prices (premi- for the same service rendered. Example: An exact copy ums) increase because insurance is based on the prin- of the original claim is not filed for the second time, ciple of solidarity, hence the loss is distributed to all but rather some portion like date is changed to get the participants in the insurance relationship and, as the benefit twice the original. In this approach, first, the loss increases, the contribution of each participant to insurance claims are clustered according to the disease cover them increases. type using Evolving Clustering Method and then they
Identification of insurance fraud is rather compli- are classified to detect duplicate claims using Support
cated. Such type of fraud can be practiced by very di- Vector Machine.
verse people, i.e. independent of education and pro- Detection of fraudulent health insurance claims by
fessions. Most risk management solutions use rules identifying correlation or association between some of
to detect potential abuse, and some solutions seem to the attributes on the claim documents is analyzed in
learn from the examples, but as the patterns of abuse [3]. Unsupervised learning based clustering were used
change, those solutions stop work for novel types of to group health insurance claims, and then
unsuperfraud. Four models are provided for identifying po- vised association to identify the correlation between
tential health insurance abuse using AI, mostly, look- attributes, and afterwards classifiers to identify
frauding for an anomalous behavior. Which are not filtered ulent claims.
by VMD or radar tracker methods respectively. Anomaly detection is studied in [4], where
statistical decision rules and k-means clustering were applied
for historical claim data, using outliers detection and
IVUS 2020: Information Society and University Studies, 23 April 2020, association rule-based mining with Gaussian
distribuKTU Santaka Valley, Kaunas, Lithuania tion. Such outliers often correspond to fraud insurance
m"orniimkaa.nztdea.knuavniiccikuatiet@e@bpbptit.lit.lt( M(R..ZKduanniacvkiačiitu¯ėt)e;̇); claims in the data.
tomas.krilavicius@bpti.eu (T. Krilavičius) Multilayer perceptron neural network (MLP) model
and fraud diamond theory (FDT)’s fraud elements as
fraud indicators, were proposed in [
Decision tree is a classifier, which is based on the idea of identifying data set division points (branching points) [6]. The structure of the decision tree is illustrated in Fig 1.
Concepts describing the decision tree [6]: 1. Root Node is a starting point i a decision tree. 2. Splitting is a process of dividing a node into 3. Decision Node is a sub-node which splits into two or more sub-nodes.
further sub-nodes. prediction model was developed to check whether a for making the binary splits. For a feature space of size claim presented by a customer is fraudulent or non- , a subset of ℝ , the space is divided into
regions
, if a region includes data that is mostly from a single class then the Gini Index value will be small: in region that belong to class .
Ensemble learning methods are based on the hypothesis that combining multiple models together can often produce a much more powerful model [7, 8, 9]. Then, the idea of ensemble methods is to try reducing bias and/or variance of such weak learners by combining several of them together in order to create a strong not split and specifies the output result.
formances. 5. Pruning is a process of removing sub-nodes from
a decision node.The opposite of pruning is split- 3.2.1. Bagging 4. Terminal Node or a Leaf is a node that does learner (or ensemble model) that achieves better perting.
tree.
is divided into sub-nodes.
Recursive binary splitting is applied to grow a classification tree.
Gini Index can be used as a criterion One of the most popular parallel methods is Bagging (Fig. 2) that goal at producing an ensemble model that is more powerful than individual models composing it [8].
Bootstrapping is statistical approach consists in generating sample of size called bootstrap sample from placement observations [8]. Bootstrap samples can be considered as representative and independent samples of the true data distribution [8].
by randomly taking with re (⋅) = arg max [ ( | (⋅) = )] .
(4) denoted [8]:
Assuming that we have bootstrap samples of size 3.2.2. Random Forest { 11, 21, … , 1 }, { 12, 22, … , 2 }, … … , { 1 , 2 , … , },
where is
[8]:: sample, we can fit almost independent weak learners − ℎ observation of the − ℎ bootstrap
1(⋅), 2(⋅), … , (⋅) and then combine them into averaging process in order to get an ensemble model with a lower variance.
Random Forest is a machine learning algorithm that creates groups of decision trees during the learning (2) process [6]. The basic idea of random forest is that the classifier is formed by combining many binary decision trees constructed using diferent subsets of data from the original data set and randomly selected subsets of attributes. This is the main diference between random forests and bagging. The structure of the random forest is illustrated in Fig 3. 3.2.3. Boosting Boosting is sequential method that based on fitting sequentially multiple weak learners in an adaptive manner: each model in the sequence is fitted giving more ( , (⋅)) = arg min ( −1(⋅) + × (⋅)) =
, (⋅) =1 = arg min ∑ ( , −1( ) + × ( )) , weak learners [8]: Adaptive boosting updates the weights attached to each of the training data set observations.
Ensemble model is defined as a weighted sum of =1 (⋅) = ∑ × (⋅) , efect to observations in the data set that were poorly handled by the previous models in the sequence [8]. (⋅ , ⋅) is the loss/error function. Thus, instead of optiwhere (⋅) is the fitting error of the given model and mising “globally” over all the imate the optimum by optimising “locally” building and adding the weak learners to the strong model one
models, we approxby one. ers. the ( ) where ’s are coeficients and
’s are weak learn- The data contains three part: the insurance policy, the Solving optimisation problem we define recurrently
(⋅) = −1(⋅) + × (⋅) , ifts the best the training data and that is the best poswhere and
are chosen that is the model that total number of entries is sible improvement over ( −1) . Then we denote [8]:
claim application and the details of the risk assessment. Whether the claim for reimbursement is satisifed according to the terms and conditions of the insurance contract is specified by the specific binary vari(6) able. Payment Claim Period 01/01/2018 - 23/12/2019,
It was decided not to analyze the medical records in this stage of research, which eliminated all cases where the document type was an invoice for the medication purchased. Records that did not specify a risk type or service were also removed. (7)
Following the expert assessment, 27 the most important variables were selected from the data set and were used for further data analysis: 5. Experimental setup 1. SERV_INST_CITY - service institution city, Fraud detection in health insurance data is formulated 2. INSURER_TYPE – insurer status as classification task. The aim of the model is to learn (individual/juridical), classifying data in to Insurance and Pretension records. 3. SELL_UNIT_CODE – product code, 4. INS_PERS_GNDR - insured person gender, 5.1. Evaluation Metrics 5. INS_PERS_CNTR - insured person center, 6. REC_TYPE – receiver status The following evaluation metrics were used to evalu(individual/juridical), ate the performance of potential fraud detection mod7. REC_CNTR - receiver country, els: accuracy, error rate, sensitivity, specificity [10]. 8. TYPE – type of medical document, fusTiohnesme amtreiaxs,uwrehsiccha nisbaetcaabllceuwlaittehdtbwaoserdowons tahnedctowno9. AUCH_CH_ID – insurer type ID, columns that reports the number of false positives, false 10. AUCH_CH_CODE – insurer type code, negatives, true positives, and true negatives (Fig. 5). 11. DOC_CODE – type of document The confusion matrix itself is relatively simple to (original/scanned), understand, but the related terminology can be con12. ODE_IMP_REC – online data exchange feature, fusing [10]: 13. SERV_INST_PRICE – price from service institu
tion price list, 1. TP and TN indicate that the fraud and insured 14. MED_INST_PRICE – price from medical institu- events are correctly classified (predicted).
tion price list, 2. FP means that the insured event was misclassi15. REAL_SUM – amount of paid money, ifed as a fraud event. 16. DISCOUNT – amount of discount, 3. FN indicates that the fraud event was misclassiifed as an insured event. 17. RCPT_SUM – amount of money stated in the
document, Accuracy is the proportion of true results (both true 18. EV_COUNT – quantity of services, positives and true negatives) and total number of cases 19. RISK_LIMIT – maximum sum insured, [10, 11]: 20. GRP_LIMIT – maximum sum of insurance risk
group, = + . (8) 21. SUPER_GRP_LIMIT – maximum sum of insur- + + + ance super group, Error rate is the proportion of false results (both 22. INS_BIRTH - insured person birth year, false positives and false negatives) and the total number of cases [10, 12]: 23. APPLIED_SUM – amount of money presented 24. fDoIrFFp_aPyOmLe_nItN,D – diference in days between pol- = + ++ + . (9) icy start date and indemnity submission date, Sensitivity of a classifier is the ratio between cor25. SERVICE – probability that service is non-insu- rectly identified positives and actual positives [10]: rance, 2276.. SRTISAKTU–S r–iskregsruolutpo,f risk assessment = . (10) (insurance/pretension).
+
Specificity of a classifier is the ratio between correctly classified negatives and actual negatives [10]:
A variable Status is data label, if Status=Pretension data records is considered fraud event. If Status=Insurance data record is normal. After data cleaning 1997076 records left in data set, and 75069 of these records have status Pretension. In this case unintentional human mistake is also considered fraud event. It means that Pretension label reason can be "Not insured for risk", "Invalid indemnity detail", "Person not insured", "Amount is exceeded". Data preprocessing was performed with the following steps:
ical binary expression using one hot encoding technique. 4. Min-max normalization was used for numeric variables.
1. The original data set was imbalance. For dividing into training and testing data sets all existing 75069 Pretension and 100000 Insurance records were selected with the widest possible combination of variable values. 70 % of the 175069 records were assigned to the train data set and 30 % to the test data set: The confusion matrices of the results obtained by the models, where the positive class is Insurance and the negative class is Pretension, are shown in figures 6, 7, 8, 9. Results show that decision trees outperform other methods, but does not work well with Pretensions, where Boosting is performs best.
Experiments results are provided in table 1. Bagging method is the most accurate (93.87%). Sensitivity shows how well positive are predicted (in this case – normal insurance event), most accurately 99.46 % predicted by decision tree. Specificity shows how well negatives are predicted (in this case – potential fraud), the best performing method (92.91%) is boosting. According to the error rate, the best performing method • 122548 records: 70000 Insurance is bagging (6.13 %).
and 52548 Pretension. • 1 52521 records: 30000 Insurance
and 22521 Pretension.
data set were assigned to 2 data set. This data set was used to verify the efectiveness of the the best created model.
Experiments with the health insurance data set show that:
We thank Virginijus Jakštys and UAB Data house 1 data are efective for potential fraud (including for cooperation and useful insights. Research was parerroneous claims) detection. tially funded by Lithuanian Business Support Agency 2. According to the accuracy, the best performing (J05-LVPA-K-02-0013).
method is bagging (93.87 %). 3. Decision trees outperform other methods by sensitivity (99.46 %), but does not work well with Pretensions (specificity is 58.98 %), where Boost- 1(www.data-house.lt) //doi.org/10.1016/j.aci.2018.08.003. [11] F. Beritelli, G. Capizzi, G. Lo Sciuto, C. Napoli, [1] M. S. Anbarasi, S. Dhivya, Fraud detection us- F. Scaglione, Rainfall estimation based on the ining outlier predictor in health insurance data, tensity of the received signal in a lte/4g mobile in: 2017 International Conference on Informa- terminal by using a probabilistic neural network, tion Communication and Embedded Systems IEEE Access 6 (2018) 30865–30873. (ICICES), 2017, pp. 1–6. doi:10.1109/ICICES. [12] F. Beritelli, G. Capizzi, G. Lo Sciuto, C. Napoli, 2017.8070750. M. Woźniak, A novel training method to preserve [2] V. Rawte, A. Srinivas, Fraud detection in health generalization of rbpnn classifiers applied to ecg insurance using data mining techniques, 2015, signals diagnosis, Neural Networks 108 (2018) pp. 1–5. doi:10.1109/ICCICT.2015.7045689. 331–338. [3] S. Kareem, R. Ahmad, A. Sarlan, Framework for the identification of fraudulent health insurance claims using association rule mining, volume 2018-January, 2018, pp. 99– 104. URL: https://www.scopus.com/inward/ record.uri?eid=2-s2.0-85047435296&doi=10. 1109%2fICBDAA.2017.8284114&partnerID= 40&md5=32f134dd6a775decec634b9c90eb8b70.
doi:10.1109/ICBDAA.2017.8284114. [4] A. Verma, A. Taneja, A. Arora, Fraud detection and frequent pattern matching in insurance claims using data mining techniques, in: 2017 Tenth International Conference on Contemporary Computing (IC3), IEEE Computer Society, Los Alamitos, CA, USA, 2017, pp. 1–7. URL: https: //doi.ieeecomputersociety.org/10.1109/IC3.2017.
8284299. doi:10.1109/IC3.2017.8284299.
[
JHMN/2018-348. [6] J. Le, Decision trees in r, DataCamp (2018). https://www.datacamp.com/community/ tutorials/decision-trees-R. [7] G. Lo Sciuto, S. Russo, C. Napoli, A cloud-based lfexible solution for psychometric tests validation, administration and evaluation, in: CEUR Workshop Proceedings, volume 2468, 2019, pp.
16–21. [8] J. Rocca, Ensemble methods: bagging, boosting and stacking, Towards Data Sci. (2019). https: //towardsdatascience.com/ensemble-methods-\ -bagging-boosting-and-stacking-c9214a10a205. [9] F. Bonanno, G. Capizzi, G. Sciuto, C. Napoli,
Wavelet recurrent neural network with semiparametric input data preprocessing for microwind power forecasting in integrated generation systems, 2015, pp. 602–609. [10] A. Tharwat, Classification assessment methods,
Applied Computing and Informatics (2018).
URL: http://www.sciencedirect.com/science/ article/pii/S2210832718301546. doi:https: