Session 74 PD, Innovative Uses of Risk Adjustment. Moderator: Joan C. Barrett, FSA, MAAA

Transcription

1 Session 74 PD, Innovative Uses of Risk Adjustment Moderator: Joan C. Barrett, FSA, MAAA Presenters: Jill S. Herbold, FSA, MAAA Robert Anders Larson, FSA, MAAA Erica Rode, ASA, MAAA SOA Antitrust Disclaimer SOA Presentation Disclaimer

2 Innovative Uses of Risk Adjustment Jill S. Herbold, FSA, MAAA Principal and Consulting Actuary Anders Larson, FSA, MAAA Consulting Actuary Erica S. Rode, PhD, ASA, MAAA Associate Actuary June 13, 2017

3 Agenda Moderator introduction Case study 1: PCP Panel Size Background Technical challenges and approach Key learnings Case study 2: Sub-capitation for Medicare Advantage beneficiaries Background Technical challenges and approach Key learnings Conclusion and questions 2

4 Limitations The views expressed in this presentation are those of the presenters, and not those of Milliman. Nothing in this presentation is intended to represent a professional opinion or be an interpretation of actuarial standards of practice. 3

5 Case Study 1: Business Problems PCP Capacity Rules for Open vs Closed Panel Incentive Compensation 4

6 Project Solutions/Results Algorithm to predict primary care utilization Risk adjusted patients by physician practice Estimated new patient capacity by physician practice 5

7 Constraints of Algorithm to Predict PCP Office Visit Utilization Intuitive and explainable to physicians Simple to implement and update Stability in predictions by patient 6

8 Data available Billing data 3 years Primary Care Visits Specialist Visits Inpatient admissions and emergency room visits Primary care panel 2 reports, 1 year apart Listing of all patients currently considered active by the primary care network Demographic information Primary payer New patient flag 7

9 Creating appropriate training data - Definitions Training data set: Set of records used to calibrate predictive model and determine relationships between characteristics (features) and outcome (response). Data is divided into two subsets, often based on a time or date variable. Feature: Data used to gather characteristics to be used as features in model (e.g., clinical conditions, historical costs, and historical utilization of services). Data generally comes from an earlier time period than the response set described below. Response: Data used to observe the outcome you are looking to predict. Data generally comes from a later time period than the feature set described above. Prediction data set: Set of records used to make new predictions using the model calibrated on the training data set. This data set is similar to the training feature data, but likely for a more recent time period (often the most recent time period). It is used to gather the same characteristics as the feature portion of the training data above. 8

10 Creating appropriate training data - Challenges Billing data had no corresponding eligibility listing Snapshots of patient panel report could be used to limit the patients Recent utilization led to inclusion in patient panel Predictions on new data would only be for patients on current panel All patients on panel at January 2016 had at least one office visit between January 2014 and December Using this same list of patients to train the model and to make new predictions would create bias. Prediction Feature Period Prediction Response Period Training Feature Period Training Response Period 1/2014 7/2014 1/2015 7/2015 1/2016 7/2016 9

11 Selecting predictive algorithm - Considerations Option 1 Generalized Linear Model Well known and fairly easy to understand Relationships are interpretable Modest run-time Does not handle interactions and multicollinearity well Linear relationships not necessarily appropriate (e.g., age) Option 2 Decision-tree Based Model Less familiar to many clients Relationships between features not as easy to interpret Longer run-time Handles interactions and multicollinearity Captures non-linearity 10

12 Selecting predictive algorithm - Accuracy Generalized Linear Model (GLM) Gradient Boosting Machine (GBM) 11

13 Balancing interpretability with accuracy Limited the number of clinical conditions used as features Focused on common conditions Conditions with intuitive relationships to physicians Balance with desire for greatest accuracy Final conditions Diabetes, asthma/copd, heart arrhythmias, CHF, pulmonary embolism, major depressive disorders, unstable angina, respiratory arrest Also included simple count of all HCC 1 conditions Error Metric Mean Absolute Percentage Error 30 Conditions + HCC Count 15 Conditions + HCC Count 8 Conditions + HCC Count HCC Count Only No Conditions 10.46% 10.47% 10.46% 10.61% 11.07% 1 HCC = Hierarchical Condition Category. These are groups of related clinical conditions identified using ICD diagnosis codes. Categories defined by CMS and are used in Medicare Advantage and ACA Exchange markets. 12

14 Interpreting relationships and feature importance Evaluate univariate or bivariate relationships on actual predictions Does not isolate effect of each feature, but shows general relationship Easy to understand Chart below shows relative predicted PCP visits (average set to 1.00) for patients with various counts of conditions and unique providers visited Count of In-Network PCPs and Specialists Visited In Feature Period Count of Clinical Conditions

15 Illustrating interactions/non-linearity For person with 1 PCP/spec visit and 0 conditions, marginal value of adding one condition = ( ) = 0.45 For person with 6+ PCP/spec visit and 2 conditions, marginal value of adding one condition = ( ) = 0.19 Count of In-Network PCPs and Specialists Visited In Feature Period Count of Clinical Conditions

16 Illustrating interactions/non-linearity Chart below shows the marginal value of one additional clinical condition at each point on the previous chart Value of additional conditions generally decrease with each additional condition (top-to-bottom) Value also decreases for patients with more providers (left-to-right) Count of In-Network PCPs and Specialists Visited In Feature Period Count of Clinical Conditions

17 Risk Adjusted Patients by Physician Practice Minutes for short and long OVs Number of short and long OVs by patient Average OV minutes per patient Risk-Adjusted patients by PCP/practice 16

18 Incentive Compensation PMPM per risk- adjusted patient, PCP OV risk adjustment PMPM per total cost Fee-forservice X PMPM X per X Patient risk-adjusted patient 17

19 PCP Capacity: Predicted Hours Relative to Scheduled Hours Number of Practices % 60-70% 70-80% 80-90% % 100%-110% 110%+ Estimated new patient capacity by physician and practice 18

20 PCP Capacity: Rules for Open vs Closed Panel Physician discretion Appointment times Nurse practitioners Monthly fluctuations 19

21 Case Study 2: The Business Problem A Medicare Advantage MCO who has historically transferred risk to contracting providers through global capitation arrangements. They have been using CMS-HCC risk scores to adjust capitation payments to providers who are taking global risk. They ve been approached by new provider groups who want to take risk, but only for a subset of services. How can they use risk scores to adjust capitation payments to newer partners taking only partial risk? 20

22 Medicare Covered and Non-Covered Services Medicare covered = covered under FFS Medicare Supplemental = non-covered under FFS Medicare Examples: Dental, non-emergency transport Plans compete to offer the best benefits Paid for by supplemental member premium and savings on Medicare covered benefits 21

23 Graphic depictions of bid forms ahead that may be disturbing to some Medicare actuaries 22

24 Key Bid Values II. Benchmark and Bid Development Total 1. Member Months (Section VI) 114, Standardized A/B Benchmark 1.000) $ Medicare Secondary Payer Adjustment 0.27% 4. Weighted Avg Risk Factor Conversion Factor Plan A/B Benchmark $1, Plan A/B Bid $ Standardized A/B Bid 1.000) $ Cost of FFS benefits, paid to MCO if less than bid. CMS-HCC Risk Score Benchmark x Risk Score MCO s cost to provide FFS benefits, including NBE and gain Paid to MCO if less than risk-adjusted benchmark III. Savings/Basic Member Premium Development 1. Savings $ Rebate $ Basic Member Premium $0.00 *All values in this presentation are illustrative Risk-adjusted benchmark - Bid Paid to MAO for supplemental benefits (percentage of savings) 23

25 A Global Cap Approach Expense Medicare Medicare Category Covered Non-Covered Total Medical Expenses $ $78.50 $ Admin $38.24 $5.01 $43.25 Margin $29.96 $3.93 $33.89 Total $ $87.44 $ Paid to Providers as Capitation 24

26 A Global Cap Approach Member CMS-HCC Risk Score Cost of FFS Benefits (riskadjusted) Cost of Supplemental Benefits Total Cost A 1.25 $ $78.50 $ B 0.90 $ $78.50 $ C 1.30 $ $78.50 $ D 2.20 $1, $78.50 $1, E 0.60 $ $78.50 $ Total 1.25 $ $78.50 $

27 Recall: The Business Problem A Medicare Advantage MCO who has historically transferred risk to contracting providers through global capitation arrangements. They have been using CMS-HCC risk scores to adjust capitation payments to providers who are taking global risk. They ve been approached by new provider groups who want to take risk, but only for a subset of services. How can they use risk scores to adjust capitation payments to newer partners taking only partial risk? 26

28 Milliman Advanced Risk Adjusters TM (MARA) background Several different types of models Prospective and concurrent Based on diagnosis data, drug data, or both Commercial and Medicare (diagnosis only) Outputs total risk score and service category scores Inpatient Outpatient Physician Emergency Other Prescription Drug (commercial only) 27

29 Score Interpretation Average in service category is proportion of allowed in category for calibration sample Commercial Medicare Service Category Concurrent Prospective Concurrent Prospective Inpatient Physician Outpatient Other Emergency Prescription Drugs N/A N/A Total

30 Example Female Age 73 Total Risk Score Inpatient Emergency Room Outpatient Physician Other

31 Uses of Service Category Scores Risk adjusted trends Care management Predicting IP admits/er visits Revenue Allocation 30

32 Service Category Revenue Allocation Calculate MA revenue for each member month Carve out admin and gain Split claim costs between Medicare covered and non-covered Medicare covered claim costs are aggregated across members and reallocated between providers and service categories, using MARA scores Claim costs for Medicare non-covered are not risk adjusted and added to the risk adjusted amounts. 31

33 Key Bid Values II. Development of Projected Revenue Requirement Medicare Covered A/B Mand Suppl (MS) Benefits Net Net PMPM for Reduction of MARA Service Category PMPM Add'l Svcs. A/B Cost Sh. Total Category a. Inpatient Facility $ $1.50 $12.00 $13.50 IP b. Skilled Nursing Facility IP c. Home Health Other d. Ambulance (2.00) (2.00) Other e. DME/Prosthetics/Diabetes Other f. OP Facility - Emergency ER g. OP Facility - Surgery OP h. OP Facility - Other OP i. Professional Phy j. Part B Rx OP k. Other Medicare Part B Other l. Transportation (Non-Covered) Other m. Dental (Non-Covered) Other n. Vision (Non-Covered) Other o. Hearing (Non-Covered) Other p. Suppl. Ben. Chpt 4 (Non-Covered) Other q. Other Non-Covered Other r. ESRD Other s Other t. COB/Subrg. (outside claim system) Allocate u. Total Medical Expenses $ $17.00 $61.50 $

34 Capitation by Service Category MARA Medicare Medicare Category Type Covered Non-Covered Total Inpatient Facility $ $15.50 $ Outpatient Facility $ $17.00 $ Emergency Facility $12.00 $1.00 $13.00 Physician Professional $ $30.00 $ Other Various $43.45 $15.00 $58.45 Subtotal $ $78.50 $ Admin NBE $38.24 $5.01 $43.25 Margin Gain/(Loss) $29.96 $3.93 $33.89 Total $ $87.44 $

35 Provider Groups Group Services Covered MARA Risk Scores Member Months Total IP OP ER PHY Other X Facility 50, Y Non-Facility 50, Z All 64, Total 114, Revenue Allocated (PMPM) FFS $ $ $ $12.00 $ $43.45 Supp $78.50 $15.50 $17.00 $1.00 $30.00 $15.00 Total $ $ $ $13.00 $ $

36 Revenue Allocation Group Services Covered Revenue Allocated Member Months Total IP OP ER PHY Other X Facility 50,000 $ $ $ $13.00 Y Non-Facility 50,000 $ $ $81.26 Z All 64,000 $ $ $ $13.00 $ $40.63 Total 114,000 $ $ $ $13.00 $ $58.45 Relative Risk Score/Revenue Group Services Covered Member Months Total IP OP ER PHY Other X Facility 50, Y Non-Facility 50, Z All 64,

37 Questions for discussion 36