Session 84 PD, Predictive Analytics for Actuaries: A look At Case Studies in Healthcare Analytics. Moderator: Carol J.

Transcription

1 Session 84 PD, Predictive Analytics for Actuaries: A look At Case Studies in Healthcare Analytics Moderator: Carol J. McCall, FSA, MAAA Presenters: Lillian Louise Dittrick, FSA, MAAA Wu-Chyuan (Gary) Gau, ASA, MAAA, Ph.D. Sheamus Kee Parkes, FSA, MAAA Andrew Morrison Webster, ASA, MAAA

2 Predictive Analytics for Actuaries: A Look at Case Studies in Healthcare Analytics (Session 84) LILLIAN DITTRICK, FSA, MAAA DR. WU-CHYUAN (GARY) GAU, PHD, ASA, MAAA SHEA PARKES, FSA, MAAA ANDREW M. WEBSTER, MS, ASA, MAAA June 16, 2016

3 Predictive Analytics for Actuaries Lillian Dittrick, FSA, MAAA Strategic Analytics

4 UnityPoint Health System Profile

5 UnityPoint Health System Map

6 UnityPoint Health Analytics PREDICTIVE ANALYTICS USE CASES

7 UnityPoint Health Analytics Predictive Analytics Use Cases Risk Stratification Blood Management LOS Prediction Clinic Appointment No Show Model Natural Language Processor Chronic Conditions with Unrecorded Diagnosis Codes Impacts Care Management Risk Coding Risk Stratification Social Determinants

8 UnityPoint Health Analytics Risk Stratification Prior 12 Months Subsequent 12 Months Training Data Outcome Regression Model Strategy 2 Year Time Horizon 100+ Total Predictor Variables Demographics Specific Diagnoses Utilization of health services

9 UnityPoint Health Analytics Blood Management: Physician Level Scorecard

10 UnityPoint Health Analytics ETG Savings Model Risk adjust episode cost by DRG severity factors to identify clusters of like patients to analyze Model risk adjusted cost difference using classification and regression trees to identify driving factors of high-cost episodes

11 UnityPoint Health Analytics Clinic Appointment No Show Model

12 UnityPoint Health Analytics Natural Language Processor Annotators are used to identify valuable facts in unstructured (natural language) notes sections of the electronic medical record. Entire data sets can be analyzed for data mining. 11

13 UnityPoint Health Analytics Natural Language Processor Annotators are used to identify valuable facts in unstructured documents (e.g. clinician notes, consult reports, free text fields in electronic medical record) and convert to a structured form 12

14 UnityPoint Health Analytics Natural Language Processor Annotators for Diabetes and COPD Testing on Calendar year data to identify coding opportunities Patient Care CMS HCC risk score Risk Stratification Other Chronic Diseases Social Determinants

15 UnityPoint Health Analytics LESSONS LEARNED

16 UnityPoint Health Analytics What Worked and What Didn t Must haves for use case identification, definition, creation, adoption and enhancements Physician engagement Use case selection process Data quality considerations Method of delivery Pilot/test sites Measurement of success Feedback loop Process for enhancements

17 UnityPoint Health Analytics Goal: Alignment of Analytics, Best Practice, and Adoption What is needed? How can we support? What happens if you don t have all three? 16

18 UnityPoint Health Analytics ANALYTICS Analytics Alone If we build it they will come. No real outcomes improve. ADOPTION BEST PRACTICES

19 UnityPoint Health Analytics ANALYTICS Best Practice Alone ADOPTION Academic ideas with no practical application. No real outcomes improve. BEST PRACTICES

20 UnityPoint Health Analytics ANALYTICS Adoption/Change Alone Most clinicians disengage if best practice and analytics are both missing. ADOPTION No real outcomes improve. BEST PRACTICES

21 UnityPoint Health Analytics ANALYTICS How are we doing? ADOPTION How do we transform? BEST PRACTICES What should we be doing? You need three systems to succeed OUTCOMES IMPROVEMENT Best Practice = What should you be doing? Analytics = How are you doing? Adoption = How do you accelerate change? 20

22 Predictive Modeling in Post-Reform Marketplace Wu-Chyuan (Gary) Gau, PhD, ASA, MAAA 6/16/2016

23 What is Predictive Analytics? 22

24 Advanced Analytics Gartner defines Advanced Analytics as: the analysis of all kinds of data using sophisticated quantitative methods (for example, statistics, descriptive and predictive data mining, simulation and optimization) to produce insights that traditional approaches to business intelligence (BI) such as query and reporting are unlikely to discover. * * 23

25 Regression analysis; Generalized linear models; Time series analysis; and Decision tree analysis. 24

26 Precision Targeting The principle aim of predictive modeling is generalization. the ability to predict the future outcome on novel cases. Summarization => Personalization Identify High Risk Patients => Predict Change of Risk In contrast, the principle aim of traditional statistical analysis is inference. Confidence interval Hypothesis tests P-values, etc. 25

27 Problems with Predictive Analytics* Where Are My Actionable Insights? Software X is a black box. I put my data, and it gives me some sort of risk scores. I know that high risk scores are bad. So, what should I do next? I purchased Software Y, and it gives me a report that there have been thirty preventable readmissions in the last month. But I want to know what to do to prevent them in the future When I tracked down high-risk patients that were identified by Software Z, I found out that they were already very sick and no room for improvement. I want to know a list of patients that will be at high-risk in the future, not in the past! Wait! All those software vendors said that they do predictive analytics *: Dr. Sriram Vishwanath, The University of Texas, Austin. 26

28 Problems with Predictive Analytics* It is About Predicting Future Not Summarizing Past How much will members pay & cost in the future? Data-driven traditional actuarial PMPM How do I deliver personalized care to those at risk? Predictive analytics centric personalized medicine *: Dr. Sriram Vishwanath, The University of Texas, Austin. 27

29 Predictive Modeling Methodology 1. Define the business problem. 2. Translate business problems into predictive modeling problems. 3. Select appropriate data. 4. Get to know the data. 5. Create a model set. 6. Fix problems with the data. 7. Transform the data. 8. Build models. 9. Assess models. 10. Deploy models. 11. Assess results. 28

30 Impact of Risk Modeling Loss Ratio Risk 1 Risk 2 Risk 3 Overall LBA 1 79% 73% 73% 74% LBA 2 81% 79% 75% 77% LBA 3 99% 93% 82% 87% Overall 87% 82% 77% 80% For Illustrative Purpose Only. 29

31 Ensemble Proxy Model Total % 75.5% 76.8% 82.0% 79.3% 73.0% Build Up Model % 72.7% 77.1% 79.6% 87.6% 75.6% % 73.4% 72.6% 80.2% 82.7% 75.6% % 83.1% 78.9% 81.7% 83.3% 80.8% % 81.1% 88.9% 89.5% 93.2% 91.1% Total 68.4% 76.1% 77.8% 82.8% 89.1% 80.0% For Illustrative Purpose Only. 30

32 Where We Are Going Insurer Experience Consumer Experience, Needs, Wants, Preferences 31

33 Where We Are Going Gross Margin = (Premium + Transfer + CSR + Risk Corridor* + Reinsurance*) Paid Claim Liability Transfer: risk adjustment zero sum money transfer CSR: cost-sharing reduction for Silver variants *: temporary programs For Illustrative Purpose Only. 32

34 Competing on Analytics 33

35 Predictive Analytics Case Studies Shea Parkes, FSA MAAA

36 Limitations The views expressed in this presentation are those of the presenter, 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. 35

37 36

38 37

39 38

40 39

41 40

42 Thank you Shea Parkes, FSA MAAA

43 SOA Provider-Focused Predictive Analytics Case Study 6/16/16

44 Agenda Domain expertise/background Analytics task (What) Purpose (Why) Analytics Workflow (How) Results Conclusion

45 Accountable Care Organization (ACO) Business structure to enforce providers to be accountable for patient health An ACO is Limited Liability Company (LLC) consisting of a subset of National Provider Identifiers (NPIs) within a Preferred Provider Organization (PPO) Insurer benefits include: Claims cost reduction Increased quality of care in insurance product Release of reserve capital by transferring risk to providers Provider benefit ability to improve patient care

46 Attribution and Network Leakage

47 Implications of Network Leakage Less clinical control over patient Revenue Leakage Less revenue to ACO providers Data Leakage Loss in risk score due to loss in diagnosis capture Loss in patient data used for patient risk stratification Network leakage threatens ACO success

48 Potential Causes of Network Leakage Patient behavior Switchers Provider behavior Referral patterns Systematic ACO network inadequacy

49 Analytics Task Predict switchers apply model to new patients Predict harmful provider referral patterns Determine which effect is more impactful in managing network leakage Are patient or provider effects more significant?

50 Purpose of Analysis Guide efficient use of scarce provider resources Target switchers for patient education Target physicians for intervention Inform ACO network development efforts

51 Analytics Workflow 1. Define the problem Data collection Data cleansing Feature engineering Data engineer 2. Design Solution Exploratory data analysis Data visualization Model development Actuary / Data Scientist 3. Monitor Results Model validation Sustainable deployment Software integration Application engineer Analytics practiced within the actuarial control cycle* *Bellis, Clare, Lyon, Richard, Klugman, Stuart and Shepherd, John. (2nd Ed) Understanding Actuarial Management: The Actuarial Control Cycle. Institute of Schaumburg, IL. Actuaries of Australia and the Society of Actuaries (2010).

52 Data Engineering 1. Start with two years of historical medical claims data in Chicagoland area from a dominant insurer covering 50,000 members 2. Scrubbed to remove denied claims, out-of-network claims, and providers having only a single claim 3. Scraped NPI website to collect performing provider information 4. Define an interaction as a directed edge in referral space Initiated by family medicine or internal medicine provider 90-day subsequent medical event Interactions for <4 unique members eliminated

53 Data Visualization By Provider Taxonomy Each Node is a provider Edges exist if there is an interaction Edge color intensity indicates the number of unique members Interactive (pan, zoom, drilldown) visualization of referral space

54 Data Visualization By Provider Name

55 Data Visualization Patient/Provider View Blue nodes are providers Red nodes are patients Edge is interaction Good patients show as clusters around a blue node Splitter patients show in isolation with many edges

56 Feature Engineering Patient-level features Demographics Age Gender Relationship to subscriber Provider-level features Subsequent provider taxonomy

57 Model

58 Results Patient-level features Younger (age 18 36) male less likely to exhibit leakage Provider-level features Specialist visits have a lower chance of leakage versus hospital and lab

59 Deployment Web-based Analytics Tool Easy and sustainable model consumption by all levels of stakeholder sophistication

60 Challenges Lack of referral data causing need for intensive data engineering Incomplete data resulted in need to use web-based data collection methods Intensive computation required to process data and visualize networks

61 Thank You! Conclusions 1. Network leakage driven both by patient and provider features 2. Following the analytics workflow resulted in continuous monitoring tool for risk of network leakage for ACOs

62 Questions?

63 Thank you!