Predicting Readmission Rates

Identifying Factors That Influence Predicted 30-Day Readmission Rates for Heart Disease in States with High Excess-Readmission Ratio using Center for Medicare and Medicaid Services data

By: Umesh Singh, Nate Strong, & Adrian Hall

Datasets and Methods: CMS data

CMS has been adopting measures to reduce 30-day hospital readmission for specific diseases that high healthcare burden such as heart attack, pneumonia, COPD, hip-knee surgery, and heart failure.

Our goal was to visualize importance of heart attack/chest pain as a cause of death across states and then determine factors or healthcare measures that lower the predicted readmission rates (PRR). Improving these factors can help the states with high excess-readmission ratio (ERR), defined as predicted: expected readmission rate. States with ERR greater than 1 are of interest for implementing the timeliness and effective measures to reduced the PRR. Data sources from CMS:

• Timely and Effective Care data 
• HospitalReadmission Reduction Program data

Hypothesis and Specific Aims

• Improving the timeliness and effectiveness of care in patients with heart attack or chest pain will reduce the 30-day readmission rates and deaths.
• Our aim is to determine the factors that significantly predict readmission within 30 days of discharge (30-day Readmission).
• We also aimed to identify the states with highest death rates due to heart attack and focus on states where predicted readmission rates could be reduced.

Visualization

Story

In our storyline
We have presented our visualization as five dashboards combined into a story. The first tab describes the age-adjusted death rates (deaths/100,000 population) for all causes and categorized by states. In this dashboard, using the previous/next buttons on the top, we can view the death rates across different states for an individual year between 2010-2014.

Click here to view all five dashboards.

Data-mining the Timeliness and Effective Care Measures Data

We have drilled down on deaths due to acute myocardial infarction (AMI) and have determined statistically significant effectiveness measures for reducing 30-day predicted readmission rate (PRR) and death due to heart attacks in the USA.

• These measures (score units) include
• Aspirin on Arrival (% of patients)
• Fibrinolytic Therapy Within 30 min. of ED Arrival (% patients)
• Median Time to ECG (minutes)
• Median Time to Fibrinolysis (minutes)
• Median Time to Transfer to Another Facility for Acute Coronary Intervention (minutes)
• Primary PCI Received Within 90 min. of Hospital Arrival (% of patients)

Correlational analysis were performed between PRR and  the scores for individual measures. Regression equations automatically generated by Tableau were inspected to determine the statistical significance of dependence of PRR on the determinants (measures).

Regression line showing Predicted 30-day Readmission Rate’ being significantly associated with ‘Median Time to ECG’ (p-value=0.008).

Increasing the median time increases rate.

Significant predictors
In tab three, as shown in the description of the regression lines, the only measures that significantly predict readmission are the ‘Median Time to ECG’ and ‘Primary PCI Received Within 90 minutes of Hospital Arrival’.

Regression line showing Predicted 30-day Readmission Rate’ can be significantly reduced by increasing percentage of patients with‘Primary PCI Received Within 90 Minutes of Hospital Arrival’ (p-value=0.03).

The earlier an ECG is taken or the greater the percentage of patients receiving PCI with 90 minutes of admission, the less is the predicted 30-day readmission rate. Early ECG could identify heart-attack or chest pain patients in need of angioplasty (i.e., PCI). Thus, reducing the time (scores) for these measures are predicted from our visualization to reduce 30-day readmission.

Adjusting for Excess-Readmission Ratio
We have further explored these regression equations for states that have excess readmission ratio. The associations between predicted readmission rate and measures (e.g., Median Time to ECG, PCI within 90 min.) are more relevant for states with cumulative excess readmission ratio (ERR, defined as predicted: expected readmission rate) greater than 1.0. ERR>1 is bad and <1 good="" is="" p="">

As shown in this tab we also determined if any of these measures are affected by ED volume. (number of patients attending the ED)

From the visualization it is quite evident that the ED volume may not be a significant factor in increasing either the median time to ECG or the percentage of patients receiving PCI within 90 minutes.

Identifying States with High Predicted Readmissions Rate

In the last tab, the states with high ERR (i.e. >1) could be identified using the filter tools shown at the top, by dragging the slider towards the right or by clicking on the left numeric and typing ‘1’. This activity will filter our the states that have ERR greater 1.

Similarly we can filter the states that have high Predicted Readmission Rate and are therefore the target states where the timeliness and effective care measures needs to be reiterated. Using these filter, we see that these top priority states are MD, KY, RI, NY and NJ in that order.

Conclusion

In states with ERR>1, reducing the 30-day predicted readmission rate will reduce the ERR that will have impact on reducing 30-day readmission associated penalties imposed by CMS in its final rule for Readmission Reduction Program under the Affordable Care Act.

These states can improve the scores on the measures significantly affecting Predicted Readmission Rate that will reduce ERR to less than 1, such as:

• Reducing the median time to ECG.
• And increasing the percentage of patients treated with PCI* within 90 minutes of admission. (*PCI=percutaneous coronary intervention e.g., angioplasty)

Usefulness and Difficulties of Visualizing in Tableau

Merging two large datasets was possible just at the click of a few buttons.

Data cleaning was relatively easy compared other similar applications such MS Excel®, SAS®, and JMP® (SAS).

Exceptional visualization and data mining was possible using this merged data sets even without the knowledge and skills of programming and coding.

Filtering the data allowed drilling into the dependent variables and then adjusting for states with ERR>1 and then determining which states are of specific interest for reducing the predicted readmission rates.

There are some issues with compatibility. We were not easily able to share the working files between the team members, even if we shared packaged files. Some files worked on Macintosh computers but not on PC.