Health Insurance Market

Transcription

1 Health Insurance Market Jeremiah Reyes, Jerry Duran, Chanel Manzanillo Abstract Based on a person s Health Insurance Plan attributes, namely if it was a dental only plan, is notice required for pregnancy, what kind of market coverage it has, is medical drug deductible integrated, does it have out of country coverage, does it have out of service area coverage and does it offer wellness program, we can predict if you are Health Service Account eligible. Using data mining techniques for prediction, we have a 83.05% accuracy via Random Forest Algorithm. Keywords: Health Insurance Market; Health Insurance, Health Savings Account, Health Insurance Attributes 3. Data Mining Techniques The data mining techniques that were used are called J48 Decision Trees, Naïve Bayes, Random Forest and Random Tree. The program we used to do the data mining is WEKA. We ran the data twice and used 10-fold cross validation on the first run and percentage split traintest for the latter. We used the edited PlanAttributes.csv file that had 8 columns, namely DentalOnlyPlan, IsHSAEligible, IsNoticeRequiredForPregnancy, MarketCoverage, MedicalDrugDeductiblesIntegrated, OutOfCountryCoverage, OutOfServiceAreaCoverage, and WellnessProgramOffered. 1. Introduction Health Insurance market is something most people do not know a lot of information of. We tend to just get what is available and let it be. Even worse, people just try to obtain the cheapest health insurance without knowing what they are really losing. Health Savings Account is an attribute your health insurance can have but not everybody has one. With the data we found on Kaggle s website, we are to predict if you can be Health Savings Account eligible or not based on 7 other different attributes. 2. Data We obtained the data from They had data in CSV files as well as a SQLite DB format. The downloadable zip file containing all the data amounts to 700 MB of data. The specific CSV file that we chose to use is PlannedAttributes.csv. Uncleaned, it had 92 MB of data, 176 columns and 77,354 rows of data. We had some help from a professional who works in the medical field who knows more about this topic to trim down the columns. Together, we trimmed down the data down to 8 columns. With all those columns gone, we re down to 2.7 MB of data.

2 4. Data Mining Results Figure 1 Random Forest 10-Fold Cross Validation Figure 2 Random Forest Split Training-Test Figure 3 Random Tree 10-Fold Cross Validation

3 Figure 4 Random Tree Split Training-Test Figure 5 Naïve Bayes Split Training-Test Figure 6 Naïve Bayes 10 -Fold Cross Validation Figure 7 J48 Decision Tree Split Training Test

4 Figure 8 J48 Decision Tree 10-Fold Cross Validation

5 5. Data Mining Analysis The first Random Forest classification, using 10-fold cross validation, gave us the accuracy of 82.95%. Using Random Forest classification again, but with split, yielded a very slightly higher accuracy %, a 0.1% increase. The Random Tree classification using 10-fold cross validation gave us the same exact result as Random Forest, which is 82.95% accuracy, with all true positive, false positive, true negative and false negative being exactly the same. As expected, split gave us the same exact result again, yielding a 83.05% accuracy, with all true positive, true negative, false positive and false negative being the same. Using J48 Decision Tree this time around with 10-fold cross validation, we obtained an accuracy of 82.90%, which is very slightly lower than Random Forest or Random Tree. With J48 Decision Tree again, but with split, we obtained 82.97% accuracy. This is still slightly lower than Random Forest or Random Tree s variant. Though, it is higher than 10-fold cross validation still. Lastly, we tried using Naïve Bayes. With 10-fold cross validation, we only managed to obtain 81.89% accuracy which is the lowest by far. We also tried the split but for the first time in our runs, split yielded lower results than 10-fold cross validation, at 81.84% accuracy. Having a closer look at the Naïve Bayes results, we noticed that Naïve Bayes are merely classifying all the data as No. Since the majority of our data does end up being a No, Naïve Bayes still manages to reach high accuracy. This also explains the difference in split vs 10-fold cross validation. Since there are more data that were used as test in 10-fold cross validation, it yielded a false higher result than 70-30, even though both predicted all No s. Random Tree or Random Forest with data split between training and test data gives us the best results with 83.05% accuracy for prediction. This is not very significant difference between all of the other tests that we ran, but this is by far the best one that we have obtained. 6. Research Looking at IEEE research, we found that in other countries, they used computational intelligence techniques to model the behavior of medical reviewers. Similar to our research, they used health insurance data to create their model. They were mainly looking at people having appointments or tests being carried out but without the need for it. Other researches have looked into how health insurance are now trying to go in the trend of using Big Data and using advanced tools and technologies to utilize the data to their advantage. Though this is similar to our topic in terms of having health insurance, they focus more on using the Big Data to create models for multitude of different things such as creating better decisions in a business point of view, and customer interest, which was our topic. They also use the data to try and develop systems to diagnose a lot of other diseases earlier than we normally could. Another way health insurance workers have utilized data mining is through risk management. They used data to assess the people who needs insurance and how much they will need in the longer run. This ties in to the rate of cost of the health insurance itself. A lot of these also ties with our research with the plan attributes. The more attributes the people need, the higher rates they have, only with much more factors to consider. Some have even used data mining to experiment in health insurance reform as they use automated agents as employees. This experiment can lead us to a better health insurance by being able to reform it multiple times without much cost as to retraining staff, implementing new rules, etc. Some health insurance providers actually now use genetic testing. Using the data from genetics, they will now use these data to increase the rates of people who are more genetically prone to diseases even though they are currently perfectly normal. This provides the health insurance providers more data to increase their prediction and make it more specific. 7. What We Learned As this project went on, we learned how to preprocess data. Data cleaning is a very important process for converting illegible data into something more meaningful. After much hard work of cleaning, we learned that data sets are significantly smaller when cleaned. In hindsight, it is actually not surprising given the fact that most of these data are all purpose and we re only looking to use it for a single purpose. We learned more about data mining and the tools that we need to use for it. WEKA is a very powerful tool to handle data. We also learned more about different strategies that were not discussed in class, such as random forest and random tree. We also learned about IEEE format and the formality of research papers. This format is specifically designed to help readers get what they want with much ease to the eyes especially that it is a standardized format. We also just learned that majority of people do not have access to a Health Savings Account, as Naïve Bayes has pointed out, about 81.89% of people do not have access to it.

6 8. References [1] B. Liu, "Study on Cost Control Approaches in Medical Insurance Market--Analysis on the Basis of a Principal- Agent Relationship of Three Participants," 2010 International Conference on Management and Service Science, Wuhan, 2010, pp [2] D. K. Thara, B. G. Premasudha, V. R. Ram and R. Suma, "Impact of big data in healthcare: A survey," nd International Conference on Contemporary Computing and Informatics (IC3I), Greater Noida, India, 2016, pp [3] Dube, Ryan. "Learn SQL Or Create A Simple Database With Sqlite Database Browser". MakeUseOf. N.p., Web. 4 May [4] Flávio H.D. Araújo, André M. Santana, Pedro de A. Santos Neto, "Using machine learning to support healthcare professionals in making preauthorisation decisions", International Journal of Medical Informatics, vol. 94, pp. 1, 2016, ISSN [5] "Health Insurance Marketplace Kaggle". Kaggle.com. N.p., Web. 5 May [6] M. E. Thatcher and E. K. Clemons, "Managing the costs of informational privacy: bundling as a strategy in the individual health insurance market," Proceedings of the 33rd Annual Hawaii International Conference on System Sciences, 2000, pp. 10 pp.-. [7] Nguyen, Dan. "Getting Started With Sqlite Browser Public Affairs Data Journalism At Stanford University". Public Affairs Data Journalism at Stanford University. N.p., Web. 5 May [8] "Resources". Tableau Public. N.p., Web. 7 May [9] S. J. Rassenti and C. A. Johnston, "Health Insurance Reform in an Experimental Market: Human Subjects, Agents Combined to Study Complex Regulatory Reform Proposals," 2009 International Conference on Computational Science and Engineering, Vancouver, BC, 2009, pp [10] Weka Data Mining Tutorial For First Time & Beginner Users Web. 4 May 2017.