Developing an All-Patient Risk Model in a Unified Analytics Environment Eric Hixson PhD, MBA Senior Program Administrator Michael Lewis MBA Senior Director
Analytics Enablement Operational Integration Insight Creation and Consumption Analytics Engines Organized Data Simplified Accessibility Performant Scalable Infrastructures Maker Space Interconnected Ecosystem Data Integration
Analytics Enablement Ops APIs {JSON} Operational Integration <xml/> Analytic Tools Insight Creation and Consumption Leadership Staff Governance Stewardship Enterprise Data Vault SAS In-Memory Analytics Analytics Engines Organized Data Enterprise Data Views Simplified Accessibility Hadoop Performant Scalable Infrastructures Data Labs Maker Space External Data Interconnected Ecosystem Data Asset Management Data Integration Metadata Manager Data Quality Business Glossary Master Data Management
Analytics Enablement Ops APIs {JSON} Operational Integration <xml/> Analytic Tools Insight Creation and Consumption Leadership Staff Governance Stewardship Enterprise Data Vault SAS In-Memory Analytics Analytics Engines Organized Data Enterprise Data Views Simplified Accessibility Hadoop Performant Scalable Infrastructures Data Labs Maker Space External Data Interconnected Ecosystem Data Asset Management Data Integration Metadata Manager Data Quality Business Glossary Master Data Management
All-Patient Risk Model
Applying Analytics to Population Health Cleveland Clinic Primary Care Risk Model CCPC Risk Model The goal of the CCPC risk model: Develop a risk score for all patients living within our service area who have received treatment at the Cleveland Clinic in the last 3 years. AHRQ Clinical Social History HCC Clinical Social History MARA Clinical Social History Target: Predict Future Direct Cost Prior Utilization Prior Utilization Predictive variables: 300+ variables considered for the model; Diagnosis grouper logic (AHRQ), prior utilization information, clinical/lab measures, and demographic information Demographics Demographics Top Predictive Variables Days Since Last Visit PCP Affiliation/Exist Flag Used to prioritize risk registries for care coordination activities Best Model Fit: 15 different models evaluated daily Previous Years Cost Count of AHRQ Groups Discharge Status County Payer Patient Age
Preventability Applying Analytics to Population Health Identifying patients who are at risk for increased healthcare spend Most risk models will tell you that high cost one year will predict high cost the next At some point, that cost becomes unpreventable Model Used: Logistic regression Maximizing the skills of care coordinators means identifying patients who are rising risk but haven t yet fallen off the cliff Risk How can we define the rising risk population? From data mining: 10% of Cleveland Clinic s attributed Medicare ACO population had low expenditures in CY1 that tripled in CY2 75% of the cohort who had such a significant increase had an admission in CY2 In order to identify the rising risk cohort, we should be identifying patients at a high risk for preventable admissions (ambulatory care sensitive)
Applying Analytics to Population Health C-Statistic by Future Time Period Predicted Ambulatory Care Sensitive Admission Model 0.893 0.883 0.876 To identify the rising risk population, we modeled future inpatient admission risk for specific conditions which our clinicians considered impactable Target variable: binary outcome, whether or not the patient had an inpatient admission within 3 different time periods (3 separate models) ACS Admission 3- Month ACS Admission 6- Month ACS Admission 12- Month Predictive variables: 200 variables considered for the model; historical diagnosis data, prior utilization information, and clinical/lab measures Best model fit: logistic regression Top Predictive Variables Patient with CKD Asthma COPD Last egfr Diseases of the Heart Number of ED Visits Diabetes Mellitus
Net Revenue Applying Analytics to Population Health Identifying the services needed to keep patients in network & provide them with the best possible care Each year, the Cleveland Clinic is attributed patients with a wide range of expenditures and network preferences. Understanding nuances in these cohorts can help us create interventions or see network weaknesses.
Net Revenue Applying Analytics to Population Health Keepers vs. Leakers K-means clustering was used to analyze the annual expenditures and network utilization preference (leakage or keepage) of attributed Medicare beneficiaries. Only two variables were used to cluster the population: annual reimbursement and percent leakage Unique traits of the leakers : - Higher propensity to use home health services - Tended to be in the far western suburbs - Twice as many patients suffering from dementia or other cognitive diseases than any other cohort 10
Applying Analytics to Population Health Understanding why patients choose another network for some of the most critical procedures Elective Admission Model As a potential countermeasure to high-revenue services leaking out of network for our attributed Medicare population, this model sought to predict whether a patient would choose the Cleveland Clinic network for elective admissions (did not come through the ED) Target variable: binary outcome, elective admission occurred in the Cleveland Clinic network or not Predictive variables: member demographics (county, age, gender), network utilization in the prior year (office visits, overall leakage) Best Model Fit: Random Forest 11