Data driven models to minimize hospital readmissions

© 2013

Data driven models to minimize hospital readmissions

Miriam Paramore, EVP Strategy & Product Management, EmdeonDavid Talby, VP Engineering, Atigeo

Hospital Industry Subject to Hospital Readmission Penalties – Oct. 2012

“Medicare Revises Readmissions Penalties – Again,” Kaiser Health News, March 14, 2013, http://www.kaiserhealthnews.org/stories/2013/march/14/revised-readmissions-statistics-hospitals-medicare.aspx

2 million

$17.5 billion 19%

2,207

$280 million

276 hospitals

“That may not sound like a lot, but for hospitals already struggling financially—especially those serving the poor—losing 1%-3% of their Medicare reimbursements

could put them out of business.”

Hospital Industry Subject to Hospital Readmission Penalties – Oct. 2012

Model Model’s Goal Sample size ContextCharlson morbidity index (1987) 1-year mortality 607 1 hospital in NYC,

April 1984Elixhauser morbidity index (1998)

Hospital charges, length of stay & in-hospital mortality 1,779,167 438 hospitals in CA,

1992LACE index(van Walraven et al., 2010)

30-day mortality or readmission 4,812 11 hospitals in

Ontario, 2002-2006LACE index + CMGs (van Walraven et al., 2012)

30-day mortality or readmission 100,000 All hospitals in

Ontario, 2003-2009

Why are new readmissions predictive models necessary?

Medical claims > 4.7 Billion

Pharmacy claims > 1.2 Billion

Providers > 500,000

Patients > 120 million

Our dataset:

• Hospital, outpatient & physician visits• Under a single master patient index• Cross-US geographic coverage

• Infrastructure requirements– Model based on the entire dataset– Model based on continuously updating data– Experiment with & combine multiple:• Modeling techniques• Feature combinations• Ways to combine the datasets

– Data quality as an integral and critical component• Missing data, errors, fraud, outliers, flurries, …

Yes, this is a big data problem

• Tens of modeling & statistical techniques apply– Without over-fitting

• An ensemble approach applies– Combine multiple ‘weak’ models

• Automated feature engineering applies– Don’t assume features, “let the data speak”

More data = Fundamentally better prediction

LACE

New Model

0.5 0.55 0.6 0.65 0.7 0.75

C-Statistic over patients discharged for AMI, HF & PN

LACE

New Model

0 20 40 60 80 100 120 140

Number of features in model

Models must be tailored

• Do not train on one hospital / geography / specialty / patient demographic and blindly apply to others• Models must be tailored for each hospital location• Do not assume which variables are most important to change

• Locality (epidemics)• Seasonality• Changes in the hospital or population• Impact of deploying the system• Combination of all of the above

Automated feedback loop & retrain pipeline is a must

Models must continuously evolve

• Yes, this is a big data problem• More data = Fundamentally better prediction• Models must be tailored• Models must continuously evolve

Key things to remember

Readmission Analysis Shows High Heart Failure Diagnoses

Identify High Risk Patients at Registration

Identify High Risk Patients at Registration: Case 1

24 Months• 192 treatments at 12 different locations• 8 outpatient visits in 2 separate facilities• 130 outpatient diagnostic or clinic visits in 14 different

facilities• Most clinical care is rendered by a PCP internal medicine practice over 92 visits

Identify Risks in Prescription History

Follow High Risk Patients Post Discharge

Thank you!