Patient Data, Population Health, a Predictive Model and Reducing Readmissions

Predictive AnalyticsBy Sarianne Gruber
Twitter: @subtleimpact

Worthy of recognition is the paper “Predicting Readmission Risk with Institution Specific Prediction Models authored by Shipeng Yu, Alexander van Esbroeck, Faisal Farooq, Glenn Fung, Vikram Anand and Balaji Krishnapuram, a collaborative team from Siemens Healthcare and University of Michigan.  This readmission model study is illustrative as to why healthcare analytics has a vital role in meeting challenges of new policies, managing population health, pinpointing gaps in existing modeling methods and in predicting a patient’s risk to readmit. This review provides a historical, analytical perspective on a new institution-specific prediction model.

Hospitals need to Reduce Readmissions
The Hospital Readmission Reduction Program, a provision under the CMS legislation, is effectually penalizing hospitals that have excessive readmission rates. The 2015 penalty fee is now 3% of the total Medicare reimbursement, which increased from 1%. A known cost driver of healthcare, readmissions contribute to a significant proportion of total inpatient spending. Defined as an admission to a hospital within a 30 day time frame, a readmission can occur at either the same hospital or a different hospital and can involve planned or unplanned surgical or medical treatments.

Population Health
Heart Failure, Acute Myocardial Infarction and Pneumonia are the disease groups presented in the study. Within each of these disease groups, a large body of research is in force to reduce readmissions. Yu and co-authors delivered to date “the most comprehensive experimental study on hospital-specific and condition-specific readmission risk predictions”.    The team built many models at “admission time and discharge time, leveraging the different available data, and experimented on HF, AMI, PN as well as all-cause all-condition readmission risk prediction”.

Gaps in Readmission Research

  • Focusing on improving discharge process and/or care transitions.   Educating patients on follow-up care and home medications, transitioning discharge information to primary care physicians, scheduling home visits and follow-up calls track continuity of care to avert readmissions. Transitional care risk assessment stratification models are ideally and clinically relevant to those of greatest risk, triggering an early care intervention during a hospitalization.
  • Using models for the general readmission population.   The LACE model, an index score, predicts early death or an unplanned readmission post discharge, and was modeled using a 4,000 patient sample from 11 hospitals in Ontario. Its drawback is that the Canadian base readmission rate is 8% and the Medicare population has a base readmission rate around 20%. LACE score is calculated by summing the assigned points per category for each attribute for each patient. The attributes are: Length of Stay (L), Acute admission (A), Charlson Comorbidity Index (C) and visits to ER in past 6 months (E).
  • Models with limited patient variables. A literature review documented that “few studies considered variables associated with illness severity, overall health and function, and social determinants of health”. Also, recent readmission models did not include “timeliness of post-discharge and follow-up, coordination of care with primary care physician and quality of medication reconciliation”.

A New Predictive Model
The status quo leads research towards an institution-specific readmission risk model.The new approach begins with:

  • Data. Extracting past patient data from the hospital/health system including demographics, labs, medications, ICD and CPT codes, etc. Structured data only is used.
  • Same hospital readmission. It identifies patient readmissions to same hospital within 30 days of discharge.
  • Predicts readmission to same hospital. It combines all the available information for each patient and builds a statistical model to predict readmission to the same hospital.
  • Disease management. For a condition-specific risk prediction model, use patient sample of selected disease to recreate model.
  • New Patients. The model can be applied to new patients to output a readmission risk score.
  • Sample. Tested on three large hospitals.
  • Dependent variable. Test two approaches. A binary classification approach (1= patient was readmitted within 30 days, and 0 not readmitted), or a prognosis (survival) analysis problem leveraging the date between discharge and readmission.
  • Applied Analytics. Support Vector Machine, a machine language and data mining technique, for classification. Cox regression, a statistical method, for survival analysis.
  • Validation. Learned models compared to LACE technique.

The takeaways of the study:

  • Discovery. Hospital-specific risk predictive models are limited given that the most predictive variables are socioeconomic, which are not generally available on the hospital EHR system.
  • Population Health Management Matters. The providers using these models care more about the higher-risk subset, for “which they need to take action”.

Recommend reading the complete study. Please click on the link “Predicting Readmission Risk with Institution Specific Prediction Models” for the details.