Predictive Modeling for Improved Outcomes & Quality Performance March 26, 2015 Presented by: Scott Zasadil Chief Scientist, UPMC Health Plan Kim Browning, CHRS, PMP, CHC Executive Vice President, Cognisight
Part I: Predictive Modeling for Improved Outcomes & Quality Performance March 26, 2015 Presented by: Scott Zasadil, Chief Scientist
Competitive Advantage Levels of Analytics Framework What s the best that can happen? Optimization What will happen next? Predictive Modeling What if these trends continue? Forecasting Why is this happening? Statistical Analysis What actions are needed? Alerts What exactly is the problem? Query Drilldown How many, how often, where? Ad hoc Reports What happened? Standard Reports Degree of Intelligence From Tom Farre, The Analytical Competitor, in Analytics: The Art and Science of Better, ComputerWorld Technology Briefing. 3
Predictive Modeling o 10% of patients account for 70% of all healthcare expenditures Goal: Use Predictive Modeling to identify these patients o Predictive Modeling identify patterns in data Requires an Enterprise Data Warehouse (EDW) and Disease Registry o Any knowledge that is uncovered can potentially be useful An informed decision is better than random guessing 4
Predictive Modeling, cont. o Uses prior utilization to Identify health risks and their importance Forecast future medical/r x utilization Goal of risk stratification is to make comparisons between similar patient populations o Expansion of input variables due to adoption of electronic health records EHRs are highly variable o Predictive Modeling tools tend to be opaque Clinical judgment is needed to be effective A risk score by itself is not actionable 5
Predictive Modeling Applications o Care management o Premium setting o Quality monitoring and improvement o Provider performance assessment o Predict admissions, readmissions, and ED use Models are not very accurate 6
Data Sources Zip Code 7
Issues with Healthcare Data o Non-normal data o Rare outcomes o Numerous ICDs, CPTs, and DRGs Increased opportunities for data entry errors Aggregate using Clinical Classification Software (CCS) and Berenson Eggers Type of Service (BETOS) o Diagnosis may depend upon the physician o Accounting for illness severity Charlson Comorbidity Index Hierarchical Condition Categories (Medicare only) 8
Illness Severity/Case Mix Measures o Charlson Comorbidity Index (CCI) o Adjusted Clinical Groups (ACG) o Chronic Disease Score (CDS) o R x Risk o Area Deprivation Index (ADI) o Disease or Pharmacy Count o E&M Visits o From a modeling perspective, the type of illness is usually not the main driver of resource utilization 9
Admits for Medicare Patients The number of MDCs is a strong driver of admits not any specific MDC by itself Source: J. Wolff et al, Archives of Internal Medicine, 2002 10
Dangers of Using a Threshold Model: Regression to the Mean Danger of chasing the money 11
Regression to the Mean, cont. Expenditures decrease in Year 2, even without intervention 12
Optimal Goal: Find Easy to Understand Rules o Many models, e.g. Neural Networks, are black boxes o Caregivers are mostly interested in the high readmit rate predictions o Physicians prefer easy to follow sets of rules 13 HOSPITAL Readmission Model Donze, et. al., Potentially Avoidable 30-Day Hospital Readmissions in Medical Patients: Derivation and Validation of a Prediction Model, JAMA Intern Med. 2013, 632-638 Hemoglobin at discharge Oncology Sodium level at discharge Procedure during index admission Index Type of admission (elective vs. non-elective) Admissions during previous 12 months Length of stay
Find Easy to Understand Rules, cont. 14
Readmission Models o Nearly all models focus on A specific population A specific disease A single hospital 18% of all 30 day readmits occur at a different hospital o Target (30-day readmit) is a relatively short window Target is small which makes modeling a challenge o Models do not focus on actionability o Any improvement over the base readmit rate is valuable o Can be used to perform hospital profiling 15
Identification of Patients at Risk for Readmission o Allaudeen, Schnipper, et. al. Inability of Providers to Predict Unplanned Readmissions, J Gen Intern Med. 2011 Jul;26(7):771-6 159 discharges Age 65 Evaluated readmission predictions from: Attending Physician, Resident Physician, Intern Physician, Case Manager, Nurse None of the predictions differed from chance 16
Avoidable Readmissions o van Walraven, Bennett, et. al. Proportion of hospital readmissions deemed avoidable: a systematic review, CMAJ, 2011 Apr 19;183(7):E391-402 Review of 34 studies 27% of readmissions are avoidable o van Walraven, Jennings, et. al. Incidence of potentially avoidable urgent readmissions and their relation to all-cause urgent readmissions, CMAJ, 2011, Oct 4;183(14):E1067-72 4812 discharges Readmissions reviewed by at least 4 physicians 16% of the readmissions were potentially avoidable 17
Causes of Readmissions o Feigenbaum, Neuwirth, et. al. Factors Contributing to All- Cause 30-Day Readmissions, Med Care, 2012 Jul;50(7):599-605 537 Readmissions Most frequent cause was lack of referral to auxiliary services Medication error was a contributing cause in 28% of all avoidable readmissions 23% of patients identified lack of medication information as being a factor 18
Actionability / Impactibility o Focus on ambulatory care sensitive conditions Asthma, Diabetes, Heart Failure, Hypertension, etc. o Focus on gaps in care o Patient activation measure o Similarities to previous patients Can model using K Nearest Neighbors (KNN) o Modeling which patients are actionable is the most critical area that needs to be addressed by predictive models Vendor tools are too general In-house data scientists lack the required skill set 19
Part II: Predictive Modeling for Improved Outcomes & Quality Performance March 26, 2015 Presented by: Kim Browning, CHRS, PMP, CHC Executive Vice President
20 Today s Agenda Objectives Top Line Revenue Infrastructure & Operations Predictive Modeling Works o Care management services Care management risk score o Dashboard o Provider engagement o STARs impact Samples & Resources Q&A
Objectives How to impact top line revenue and manage utilization costs to improve bottom line Get beyond the rhetoric Actionable items that work 22
23 Top Line Revenue Risk adjustment impact o Persistency and suspects Persistency = elementary intervention Suspects = use of algorithms to predict likeliness of a condition The stronger the analytics, the better the results Predictive modeling isn t used to prevent services; rather to ensure conditions are captured Predictive modeling premise to anticipate member status changes o Get out in front of them ESRD Duals Disabled Aged
Infrastructure & Operations Begin with organizational infrastructure and resource support in place o o o Doesn t have to be perfect to get started Work with what you can any informed action is better than guessing Work the output! 24
Data Integration Infrastructure 25 Current state: assemble all patient data with multiple data source integration Care Connect Epic EMR (RRHS) GRIPA Connect Exchange (Infor) Practice EMRs (coming soon) Other Community Sources: o ACM Lab o EWBC o RHIO: Strong, Highland, Clifton Springs, Thompson, Geneva, Noyes, United Memorial, Borg-Ide Data Warehouse Analytics & Reports 100 + Practice Management Systems Payer Claims
26 Infrastructure: HIE View
Predictive modeling works o Work the Utilization Costs 27 Greater Rochester Independent Practice Association (GRIPA) Uses predictive modeling to provide care management services Chronic condition management Cardiac risk management Diabetes prevention Clinical R x program
Chronic condition management o Care Management Services 28 Strive for reduced ED visits and admissions timely A1c, FLP, visits for with CAD and DM Cardiac risk management o Target hyperlipidemia and HTN timely visits, kidney function testing, FLP Diabetes prevention o Target high risk members
Elements of care management risk score o 16 Chronic conditions and their associated co-morbidities o Utilization markers # ED visits # Hospital visits # Physician visits o Socio-economic marker Zip code Patients/members are scored and managed o Patient outreach report o Refreshed monthly Care Management Services, cont. 29
Budget is managed via Dashboard o Pharmacy cost savings: 50% 30 Pharmacist(s) and CMO outreach o High cost/high risk patients: 25% Top 1 2% of patients; proactively closing gaps Coordinated care network o Readmission reduction:25% Better transitions of care Care Management Services, cont. Readmission Reduction 25% High Cost/High Risk Patients 25% Pharmacy Cost Savings 50% Includes home visits with medication reconciliation and developing care plans B u d g e t
Provider Engagement Budget is managed by provider engagement o CMO review of physician performance o Clinical Integration Committee review PCPs Monitors quality and budget SCPs o Physician incentive 31 $ based on individual performance o Recruitment influence
Ensure providers know their high risk patients o o Provider Engagement, cont. 32 Providers, know your high risk patients and give them a free pass Squeeze the patient in Prevents admissions or re-admissions Mark H. Belfer, DO, FAAFP Chief Medical Officer, Cognisight & GRIPA
33 STARs Impact Care Management Provider Engagement + = Closing Gaps in Care Drive the Results STARs will follow
Samples & Resources Patient literature Patient dashboard High cost R x outreach Financial dashboard Readmission intervention programs literature references GRIPA s Value Report 34 http://www.gripa.org/documents/publications/2013%20value%20report.pdf
35 Patient Literature GripaConnect.com o o Real-time data reports at your fingertips Let s you proactively manage patient care Care management support team o Full service, multi-disciplinary team to facilitate patient health improvements GRIPA support team o Provide customized support to help physician practices optimize their patient performance and results Proprietary Resources and Population Health Management Tools Accountable Care communication support resources o o o Missing lab reports High cost drug reports Patient outreach reports
36 Patient Dashboard
37 High Cost R x Outreach Confidential and proprietary information removed
38 Financial Dashboard
Readmission Intervention Programs Literature Jack, Chetty, et. al. A Reengineered Hospital Discharge Program to Decrease Rehospitalization, Ann Intern Med, 2009 Feb 3;150(3):178-87 (Project RED) 39 o Simplified, easy to understand, patient instructions; follow-up phone calls; dedicated discharge advocate o Patients are 30% less likely to be readmitted or have an ED visit Hansen, Greenwald, et. al., Project BOOST: Effectiveness of a Multihospital Effort to Reduce Rehospitalization, J Hosp Med, 2013 Aug;8(8):421-7 o 11 hospitals o 2% absolute reduction in readmission rate
Readmission Intervention Programs Literature, cont. Coleman, Parry, et. al. The Care Transitions Intervention Results of a Randomized Control Trial, Arch Intern Med, 2006, Sep 25;166(17):1822-8 o 40 Randomized Control Trial for Age 65 and 1 of 11 specific diagnoses Naylor, Advancing High Value Transitional Care The Central Role of Nursing and Its Leadership, Nurse Admin Q, 2012, Apr-Jun;36(2):115-26
41 Q&A For more information, please contact: Kim Browning (e) kbrowning@cognisight.com (p) 585.662.4215 www.linkedin.com/in/kimbrowning1