Blue Cross & Blue Shield of Rhode Island New Approaches Focusing on Dynamic Variables Related to Changes in Member’s Health Status Suizhou Xue September 2008
2 Background and Objectives Predictive Modeling at Blue Cross & Blue Shield of Rhode Island Predictive Modeling for Underwriting: Small Group and Large Group Predictive Modeling for Case Management New Approaches on Dynamic Variables for Early Identification Predictive Modeling for Disease Management Blue Health Intelligence for Risk Analysis
3 History of Predictive Modeling at Blue Cross & Blue Shield of Rhode Island 1990s Rules based predictive models used for Case Management identification Began researching Predictive Modeling/Data Mining methodologies and software Developed Predictive Models for Case and Disease Management using statistical methods Incorporated Johns Hopkins Predictive Models output into Case and Disease Management Initiatives Beta tested Johns Hopkins Predictive Modeling Software Timeline 2006 Incorporated Predictive Models into Underwriting process 2008 Incorporated detailed Pharmacy Data into Predictive Modeling process Included Dental data in Predictive Modeling process when available 2007 Health risk appraisal data included in Predictive Modeling process Introduced dynamic variables to identify changes in member health status for earlier identification Remodeling Process for Disease Segmentation Implemented BHI Risk Score Benchmarks into Account Reporting & Analysis
4 Technology of Predictive Modeling Johns Hopkins Predictive Models based on both Diagnosis (Dx) and Pharmacy (Rx) Information Angoss KnowledgeSTUDIO Predictive Model / Data Mining Combination of software allows for customization and inclusion of claims, PHA, gaps in care, biometric and dynamic variables Blue Health Intelligence DCG Benchmark and Statistics
5 Blue Cross & Blue Shield of Rhode Island Distribution of Members by Product Type Small Account (Size < 50) 15% Medicare 11% Individual Programs 3% Large Account (Size > 50) 65% Other 6%
6 Small Group Underwriting Many Factors Involved in an Account’s Final Rates Final Rates Trends Pools Experience Admin. Expense Reserve Contribution State Regulations Medical Risk
7 Small Group Underwriting - Testing Mode Scored individual members by ACG Predictive Modeling Score (PM) and Manual Medical Underwriting Points (MU), and summarized to an account score Compared raw scores and ranking of account PM vs. MU Correlation coefficient of about.70 Created a Set of Conversion Parameters between PM and MU through regression
8 Small Group Underwriting - Implementation Implemented for 4 th quarter 2006 cycle accounts Developed supporting system for ongoing outlier review and virtual medical record access Outlier criteria includes: extreme PM values and loss ratios Medical Underwriting reviewed 15% of accounts based on criteria Only modified 3% of those reviewed Successfully delivered final score July 2006, and replaced manual medical underwriting system
9 Small Group Underwriting – System Support
10 Small Group Underwriting – System Support
11 Small Group Underwriting - Results Reduced cycle timeframe from 6 to 3 months Allows for more current claims experience Reduced Medical Underwriting Staff Improved accuracy of Medical Underwriting Improved consistency and justification of results Coordinated Corporate Predictive Modeling activities
12 Small Group Underwriting - Evaluation Actual Expense Consistent with Rating 2nd Quarter 2007 Results Related To Median
13 Large Group Population BCBSRI’s Large Group Market –385,000 Members –550 Accounts % of Accounts % of Members Account Size
14 Large Group (IER) Predictive Modeling General Process Produce electronic file with Predictive Modeling scores for each account in rating cycle Relate PM scores to specified comparable population Two comparable statistics for each account provided to underwriters –Percent difference between account’s overall PM score and the community score –Percent difference of account proportion of high risk members compared to communities’ proportion of high risk members
15 Underwriting for IER Commercial Renewals Account InformationPM (Relative to Commercial Pool) Account Number Account Name Self Funded Cycle Total Contract Total Risk Score % High Risk 4H07Tony’s IncorporatedY May %+64.91% 959Metro PropertiesNMay %-24.56% 3943Leah CosmeticsNJune % +7.02% 5V53Goldmine JewelsNJune %-14.04% Colonial GroceriesNJune % -3.51% 3129Eric Simmons, Inc.NJuly %+46.49% 1A126Michelle & Co.NJuly1, %-10.53% Califano GroupNJuly1, %-50.00% Total Quarter16, %-0.88% Predictive Modeling Claims Incurred 01/2007 – 12/2007, Paid 12/2007
16 Case Management Objectives: Identify members who are likely to be high risk/high cost in the future Drill down to explain the major components that contribute to the risk factor Intervention –Members whose health can be improved –Members who are most likely to incur future cost savings –Collaborate care
17 Case Management – PM Status Predictive Modeling Member Demographic Information Cost Distribution Predictive Modeling Risk Probability Hospital Dominant Marker Disease and Condition Profile Virtual Medical Record - By Type of Service - Chronological Case Management / Disease Management Information Quarterly Update
18 Case Management - Challenges Challenges in Predictive Modeling: Enhance model for predictive accuracy, and reduce false positive members Early identification for members whose health status could be changed in the future How can the predictive modeling program maximize its value to the case management program Actionability Timing and scope of intervention
19 Case Management – Future Health Status Prospective Member Health Status: It’s critical for Case Management to identify the members who will change health status in the future for possible early intervention Medical claims, especially pharmacy data incurred 6 months or less, instead of 12 months, were sometimes used for Case Management. It was considered that the recent claims experience was strongly associated with future health risks Generally speaking, a disease or condition is changed within a certain analysis period. Prospective expense for the coming year will be different depending on the conditions incurred in the beginning of the year and end of the year Should consider weighing the conditions incurred in different analysis periods
20 Case Management – PM Enhancement Test Predictive Modeling – Dynamic Variables Introduced dynamic variables: those variables change their values during the period of claim experiences, such as medical utilization, visits and tests. They can be expressed as their values, rankings, or moving ratios by quarter or month, for example, quarterly medical expenses and their moving ratios (4th qtr expense vs. 3rd qtr expense, etc.) Combination of ACG Predictive Modeling results, utilization, measures, and dynamic variables allow us to customize the plan data and build the enhanced predictive models: Neural Network and Decision Trees The dynamic variables, featured at the end of the claims period are displayed near the top of the splits in the Decision Tree Predictive Model. Similarly, the dynamic variables also showed the strong contribution in the Neural Network Predictive Model
21 Case Management – Predictive Modeling Decision Tree
22 Case Management – A New Approach Predictive Modeling – A New Approach The strong prediction power of the dynamic variables implies that the prediction accuracy will increase progressively from past to present medical experiences; the current claims reflect more in member’s future health status We tested three models for the latest claims for early identification: 1) ACG predictive modeling with local calibration; 2) Customized model by neural network; and 3) ACG predictive modeling Moved from quarterly, monthly, bi-weekly to weekly. The members selected for Case Management intervention are those with a probability difference of 0.7 between current weekly results and quarter base file. Implemented the weekly predictive modeling results into McKesson Disease Monitor System. The exception rule of the system makes more efficient use of the predictive modeling results
23 Case Management – System Implementation ClaimnoLineNMemberid Field break Proc code Field break2 From date Field break3 Service type BCBSRIMEMB0031||||||||||||||||| PMCMH ||||| ||||||||||||||||PM DATA BCBSRIMEMB0032||||||||||||||||| PMCMH ||||| ||||||||||||||||PM DATA BCBSRIMEMB0033||||||||||||||||| PMCMH ||||| ||||||||||||||||PM DATA BCBSRIMEMB0034||||||||||||||||| PMCMH ||||| ||||||||||||||||PM DATA BCBSRIMEMB0035||||||||||||||||| PMCMH ||||| ||||||||||||||||PM DATA BCBSRIMEMB0893||||||||||||||||| PMCMM ||||| ||||||||||||||||PM DATA BCBSRIMEMB0898||||||||||||||||| PMCMM ||||| ||||||||||||||||PM DATA BCBSRIMEMB0899||||||||||||||||| PMCMM ||||| ||||||||||||||||PM DATA BCBSRIMEMB1631||||||||||||||||| PMCML ||||| ||||||||||||||||PM DATA BCBSRIMEMB1633||||||||||||||||| PMCML ||||| ||||||||||||||||PM DATA BCBSRIMEMB1634||||||||||||||||| PMCML ||||| ||||||||||||||||PM DATA BCBSRIMEMB3938||||||||||||||||| PMCMA ||||| ||||||||||||||||PM DATA BCBSRIMEMB3943||||||||||||||||| PMCMA ||||| ||||||||||||||||PM DATA BCBSRIMEMB3944||||||||||||||||| PMCMA ||||| ||||||||||||||||PM DATA BCBSRIMEMB3945||||||||||||||||| PMCMA ||||| ||||||||||||||||PM DATA BCBSRIMEMB3946||||||||||||||||| PMCMA ||||| ||||||||||||||||PM DATA Predictive Modeling – Disease Monitor File
24 Case Management Results (Challenges) in Predictive Modeling: Enhance model for predictive accuracy, and reduce false positive members – Combined ACG predictive modeling results and other measures including dynamic variables. Decision Tree and Neural Network models increase the prediction accuracy Early identification for members whose health status could be changed in the future – Reduce time to weekly engagement in Predictive Modeling How can predictive modeling program maximize its value to case management program – Implemented the results into McKesson Disease Monitor System Timing and scope of intervention – Produced weekly member list with the highest risk scores, and grouped members in different risk tiers for weekly intervention
25 Disease Management Objectives: Identify members who are likely to be high risk/ high cost in the future within a disease segment Diabetes, Asthma, Heart Disease, Hypertension, Cancer, Depression, etc. Co-morbidity Stratification of risk score for intervention
26 Disease Management - Diabetes Medical Expense Distribution
27 Disease Management Predictive Modeling – A New Approach: The difference in expense distribution between general commercial population and specific population indicates that it’s necessary to build a new model for a disease population rather than use the model for commercial population The lack of sufficient population size prohibits us from calibrating model locally for a specific disease Combination of ACG predictive modeling results and inclusion of utilization, measures, and dynamic variables, etc. allows us to build the robust predictive model through neural network and decision trees
28 Disease Management - Results Predictive Modeling – Results: The customized model for diabetic members increases nearly 20% of predictive accuracy compared to the general predictive model for commercial population Stratification based on the predicted risk score and evaluation of co-morbidity Produce a member listing for intervention
29 Disease Management - Diabetes
30 Blue Health Intelligence – Risk Analysis DCG Risk Scores Brings together the claims experience of 79 million BCBS members nationwide Detailed DCG risk score benchmarks by geography, industry and company size BCBSRI analytical team will be actively incorporating BHI DCG risk score benchmarks into analysis and reporting
31 Questions?