Data Mining Applications in Health Insurance

Transcription

1 Data Mining Applications in Health Insurance Salford Systems Data Mining Conference New York, NY March 28-30, 2005 Lijia Guo,, PhD, ASA, MAAA University of Central Florida 1

2 Agenda Introductions to Data Mining Modeling Applications in Health Insurance Case study Discussion Guo 2

3 Introduction - Data Mining An information discovery process. Knowing your goals Understanding your data Choosing the right methods Understanding the limitations Validation and testing Make crucial business decisions Guo 3

4 Introduction DM Process Understand the Economics Define the Goal Identify Data Sources Prepare Data Transform Data Apply DM Models Validate DM Models IMPLEMENT Guo 4

5 Introduction DM Goals Identifying responsive potential customers Identifying existing customers that more likely to terminate Identifying high risk purchaser Identifying the factors that cause large claims Identifying interactions among risk factors Guo 5

6 Introduction DM Process Guo 6

7 Applications in Health Insurance Data Warehouse Underwriting Pricing/Rate Making Claim Scoring Risk Management Policy Level Analysis Variable Selection Guo 7

8 Operational Data Store Data Warehousing Example Transactions Surveys Demographics Primary Selection: WHO? Unique Patient List Transactions Surveys Demographics Pharmacy Claims Physician Claims Hospital Claims Rx Service Level Table Med Claims Surveys Group by Patient... Summary Level Table Service Level Variables Summary Level Variables Derived Variables/ Flags Secondary Selection: WHAT DATA? Tertiary Selection: WHAT DOES THE TRANSACTION DATA TELL US? Summary: WHAT DO WE KNOW ABOUT THIS PATIENT? Guo 8

9 DM in Insurance Underwriting Improving profit margin. Gaining competitive edge Risk evaluation process. Lots of variables Lots of interactions Easy to follow procedure. Decision tree can be used Guo 9

10 DM in Risk Management Reinsurance To structure more effectively by segmentation Hedging Target retention and building loyalty Guo 10

11 DM in Policy Level Analysis Surrender analysis Profitability analysis Policyholder s s behavior DM methods used Neural networks Decision trees Logistic regression Guo 11

12 Case Study - VA GMDB, GMAB, GPAF, GMIB, GMWB, Risk factors Interest rate Market volatility Surrender Timing New business Valuation Hedging the risks Guo 12

13 Case Study - VA More risk factors Equity Price Risk Mortality Risk Time to Death Volatility Fund Transfer Risk Elective Reset Risk Model Risk Guo 13

14 Case Study VA Valuation Liability estimation Asset testing Pricing volatility Cash flow testing Analysis based on Benefit levels Policyholder s investment strategy (transfer, lapse) Effects of financial variables such as inflation and interest rates Effects of changing social and political conditions Guo 14

15 Applications Variable Selection Problem -- Given {Y,X} where Find F, such that F( X) Y Find Z X, and F*, such that X = { x1, x2,... x N } F*( X) Y Improving model accuracy and efficiency Making crucial business decisions Guo 15

16 Case Study Variable Selection Group Health Insurance Line Profit margin over five year period and 0ver 100 input variables Principle Components Analysis applied 40 variables remains How to improve business profit? Guo 16

17 Function Approximation F( X) = F + β T( X) + β T ( X) β T ( X) M M F 0 is the initial guess Stegewise approximation Each stage added by reducing errors Each stage is weal linear a small tree. Sequential adjustment Guo 17

18 Example Profit=6.5% +0.8%, if AS > %, otherwise +1.2%, if male young than %, otherwise Guo 18

19 Function Approximation GIVEN Y: Output and X: Inputs or Predictors L(Y, F): Loss Function ESTIMATE F*( X) = argmin E [ L( Y, F( X))] F( X) Y, X Guo 19

20 Classical Function Approximation F = F( X, Β), Β= { β } j Β= { } β j Solve from min LY (, F( X, B)) Guo 20

21 Nonparametric Function Approximation Initial guess { F ( X )} 0 i Compute ur g L = F( X ) i N i= 1 Take a step in the steepest descent direction Guo 21

22 Gradient Boosting N 1 L({ F( X )}) = ( Y F( X )) Initial guess i i i N i = 1 { F0 ( X i )} 2 FOR m = 1 TO M g = L( F ( X )) m m 1 i Fit an L-node L regression tree to the current residuals For each given node, calculate node average residual h m ( X ) i Update: END { F ( X )} = { F ( X )} + h ( X ) m i m 1 i m i Guo 22

23 Case Study Tw o Variable Dependence for PROFIT_MARGIN; Slice SIZE = PROFIT_MARGIN Tw o Predictor Dependence For PROFIT_MARGIN Partial Dependenc REGION Tw o Variable Dependence for PROFIT_MARGIN; Slice SIZE = PROFIT_MARGIN Partial Dependenc REGION Guo 23

24 Case Study Absolute Error Absolute Error Number of Trees Train Test Guo 24

25 Case Study - Single Stats and Variable Importance Input Additive Multiplicative Importance Variable Variable Variable Variable Variable Variable Variable Variable Variable Guo 25

26 Case Study - Pair Stats and Variable Importance Variables Additive Multiplicative Variable 1 & Variable Variable 2 & Variable Variable 2 & Variable Variable 2 & Variable Variable 3 & Variable Variable 3 & Variable Variable 4 & Variable Variable 5 & Variable Variable 6 & Variable Guo 26

27 Predictive Modeling Predictive modeling using 9 variables Regression tree with multiplicative risk factors to increase predictive power Improve business profit Guo 27

28 Importance of Multiple Techniques Robust model with high predictive accuracy Practical constrains Algorithm complexity Ease of understanding of results Guo 28

29 Is Data Mining for you? Defining the goals Understanding your data Choosing the model Using multiple techniques Improving your decision making process Thank you! Guo 29