Using Predictive Analytics to Reduce COPD Readmissions
Agenda Information about PinnacleHealth Today s Environment PinnacleHealth Case Study Questions?
PinnacleHealth System Non-profit, community teaching health system Harrisburg Hospital founded in 1873 720 beds in three hospitals 43,000 discharges 114,000 ED visits Participating in CMS ACO and Bundled Payment initiatives Overall Readmission rate 11.9% Readmission rate for patients with COPD 18%, Heart Failure 25%
PinnacleHealth Vision Statistics now at the core of modern medicine Predictive analytics is now a business IMPERATIVE Hidden in the vast amounts of generated data are discoveries that could lead to better outcomes and costs Pressing need to turn data into information, information into knowledge, and knowledge into action Like many others, Pinnacle is atrophic in this space and needed to exercise the muscle; needs to become a core competency Key question is how to use this technology to improve health care delivery AND OUTCOMES versus just score keeping How to realize the business value on merit within 12-24 months
Typical approach to Predictive Analytics at Pinnacle Triggers Medical spend threshold, Inpatient and/or ER visit counts and Specific Diagnoses to identify patients Problem Identifies Patients Too Late in the Care Process to Make a Real Impact Symptom of the Problem Regression to the Mean - traditional Identification will identify some members who will naturally see a decrease in spending regardless of intervention activities
What does Pinnacle need? Appropriate Information to Identify the Patient at the Right Time and at the Point of Care Identify appropriate patients for interventions. Evaluate Risk Patients with most identifiable gaps in guidelines or forecasted acute care and assessment of cost impact. Patient specific actionable information: Clinical History Risk Profile Gap Report Disease Profiling Understanding the history Clinical Care Pathways Do standardized evidence based protocols and pathways exist?
Pinnacle s Opportunity? The Future: (Closed Loop Awareness Systems) Altering the health system s behavior in response to patient patterns in ways that will improve patient outcomes and makes the organization more successful at pursuing its goals. PinnacleHealth can identify and fulfill new clinical needs for a patient often before the patient knows themselves. To do so, new data points have to be collected to identify and simulate patient patterns. If managed properly, medical analytics can be developed to facilitate shared decision making. By modeling both patient desires and utilization of medical resources, patterns of differences will emerge. Understanding the interconnected relationships give rise to new cost effective and high-quality care models.
Pinnacle s Prescriptive Solution Build a model that is as simple as possible, yet not simpler. Provide Risk Identification and impact analysis for all patients not just catastrophic. Helps with Regression to the Mean issues as you have treatments across the Care Continuum. Forecast Days between Exacerbations and Acute Length of Stay. Understand the scalability of disease states. Individualized Action - plans per Patient. Identify and study best opportunity for achieving reductions in total costs for chronic illness care. Best opportunity to impact cost by intervening with evidence based guidelines such as Home Health and Paramedicine Interventions. Allow for Workflow Integration. Detailed patient profiles.
COPD Project Assumptions and Hypothesis Acknowledgements: There is no unique methodology to predict when a patient with COPD will be admitted/readmitted. One needs a battery of tools constructed as observables such as a registry of patients with COPD, a long treatment history, and the clinical data that is relevant to the patient dynamics as a system. Model and Intervention Testing: Hypothesis H1: Patients who are admitted with a diagnosis of COPD can be diagnosed as at risk in real time before the occurrence of a readmission for COPD. Hypothesis H2: Interventions made subsequent to the at risk prediction (new knowledge) can mitigate the likelihood of the readmission occurring.
Established Process: Integration of Analytics at Point of Care COPD COPD Analytics Methodology presented and agreed upon by Pulmonary expert 1/27/14 Medical Practice team developed bundled care pathway protocols in relation to predictive rule sets complete: standard order set provided. Research group evaluation of data requirements 2/17/14 through 3/30/14. The pulmonology team established clinical indicators and algorithms as potential predictors of COPD exacerbation and readmission. Modeling Methodology and Validation 3/01/14 to current. The modeling team translated data into machine tools that recognized COPD exacerbation patterns. Model Demonstration 4/17/14, a Live demonstration of actual patient data was presented to the Predictive Analytics Steering Committee. The committee determined that the COPD model should be trialed for a three (3) to six (6) month period for evaluation.
Integration of Analytics at Point of Care COPD continued Future Considerations - continued: Predictive Analytics Workflow Development A clinical team will create a workflow for incorporating decision support tools in existing protocols, electronic medical records (EMR) and team rounding. Create a registry to track predictions, outcomes and long-term trends Design a predictive analytic technology infrastructure for scalability. Modeling team will work with Care Management to identify further predictive rule sets around non-clinical variables Mass customize models across the health system to specific care team needs. Eventually take advantage of technology infrastructure to automate protocol execution as much as possible.
Model Criteria for COPD The Kaplan Meier analysis demonstrates the predictability of a staged disease such as COPD. Using existing medically related protocols, Pinnacle s team has developed highly accurate and reliable patient episodic predictions. We use the entire population to determine each segment of the population; then use the segments to predict what segment the patient is in.
Cardiac Comorbidity Score <= 80% Model Criteria for Readmission Risk 2 Age <69 Risk 1 or 3 Ages 69-77 Risk 1 or 2 Ages >77 Cardiac Comorbidity Score > 80% Modeling allows PH to assign risk values to every aspect of the patient episode. Algorithms developed to make decision tree model analysis reliable. General Health Score 4-80% Score > 80% Albu. and Normal Albu. or Outlier
Risk Cat 1 Model Criteria for Readmission Risk Cat. 2 Risk Cat. 3 Age < 69 Ages 69-77 Age < 77 Score 4-21% Score 21-53% Score >53% Outlier Normal Anti-Pysch. or Depress. No Anti- Pysch. Or Depress.
COPD Model Project Criteria for Readmission Summary Risk Cat 1 Cat. 2 Clinical Team developed COPD Risk Protocol: Risk Cat. 3 CAT Score Completed Smoking Cessation Consult Completed Age < 69 Ages 69-77 Age < 77 Respiratory Completed Inhaler/Respiratory Medication Education PAM Survey Completed (future) COPD Action Plan Completed by Patient and Reviewed on Rounds Patient Score Physically had all Medications Score Prior to Discharge 4-21% 21-53% Medication Teach Back Completed PCP Appointment Scheduled within 7 Days of Discharge Score >53% Pulmonary Appointment Anti-Pysch. Scheduled or within No Anti- 4 Weeks of Discharge Outlier Normal Depress. Pysch. Or Depress. PFT Scheduled for Post Discharge Para medicine Notified of Discharge and Handoff Information Given 2 visits planned Med Rec, Review Med Teach Back, Vitals, Review COPD Action Plan
COPD Modeling Criteria for Readmission Effectiveness Risk Cat 1 Risk Cat. 2 Risk Cat. 3 Difference between statistical significance and operational Outlier significance Age Operational < 69 effectiveness means that Ages 69-77 whenever Pinnacle builds a Age model, < 77 data is partitioned into three subsets: training, testing, and validation. The model is built on the data in the training subset, and then evaluated on the testing subset. Score Now Score that you ve gone through this Score process several times, however, >53% you have a 4-21% 21-53% problem because you ve been using the testing subset to help guide how you define the settings and parameters of your model for training, it s no longer fully independent. This is why you need a third Anti-Pysch. subset or of validation No Anti- data. Ideally, the data you use to test Depress. Pysch. Or your model Normal is data you ve never seen before, so Depress. once you re finally convinced that you have a good model that s consistent and accurate, you then test it against more data such as new source data (ie, administrative claims data) etc
Testing Data Model - COPD Criteria for Readmission Project Summary Risk Cat 1 Risk Cat. 2 Age < 69 Ages 69-77 Age < 77 Our Prescriptive Analytical tool can accurately identify 75% of those Outlier Score 4-21% Score 21-53% Risk Cat. 3 Pinnacle s Testing Data shows that 40% of COPD patients tend to be readmitted. patients. Also, we can predict which of those patients will be admitted, when they will be admitted, and how long they will stay. Next step: Score >53% Utilize CMS ACO administrative claims data and Commercial Payer Data in Anti-Pysch. or No Anticonjunction with Pinnacle s Depress. robust ambulatory Pysch. Or and acute medical records to Normal Depress. improve accuracy rate to at least 95%
Demonstration Risk Cat 1 Model Criteria for Readmission Risk Cat. 2 Risk Cat. 3 Age < 69 Ages 69-77 Age < 77 Score 4-21% Score 21-53% Score >53% Outlier Normal Anti-Pysch. or Depress. No Anti- Pysch. Or Depress.
Risk Cat 1 Model Criteria for Readmission Key Factors for Success Risk Cat. 2 Risk Cat. 3 Physician led from the start, kept them engaged, and they followed through. Age < 69 Ages 69-77 Age < 77 Multidisciplinary team including systems engineer, data scientist, performance and quality improvement, care coordination, primary care, and pulmonary physicians. Score Medical Score director committed resources; Score other divisions >53% (teaching, medical 4-21% 21-53% practice) followed suit. Agile development process with rapid time to delivery, bulk of model Anti-Pysch. or No Anti- Depress. Pysch. Or Outlier development Normal completed in seven weeks. Depress. Differentiated this effort from a research project.
Thank You! George Beauregard, DO SVP & Chief Clinical Officer PinnacleHealth System gbeauregard@pinnaclehealth.org