Paper 3254-2015

Predicting Readmission of Diabetic Patients using the high performance

Support Vector Machine algorithm of SAS® Enterprise Miner™
Hephzibah Munnangi, MS, Dr. Goutam Chakraborty
Oklahoma State University, Stillwater, Ok

ABSTRACT
Diabetes is a chronic condition affecting people of all ages and is prevalent in around 25.8 million people
in the U.S. The objective of this research is to predict the probability of a diabetic patient being
readmitted. The results from this research will help hospitals design a follow-up protocol to ensure that
patients having a higher re-admission probability are doing well to promote a healthy doctor-patient
relationship. The data was obtained from the Center for Machine Learning and Intelligent Systems at
University of California, Irvine. The data set contains over 100,000 instances and 55 variables such as
insulin and length of stay, etc. The data set was split into training and validation to provide an honest
assessment of models. Several variable selection techniques such as Stepwise Regression, Forward
Regression, LARS, and LASSO were used. Using LARS, prominent factors were identified in determining
the patient readmission rate. Numerous predictive models were built: Decision Tree, Logistic Regression,
Gradient Boosting, MBR, SVM, and others. The model comparison algorithm in SAS® Enterprise Miner
13.1 identified that the High Performance Support Vector Machine outperformed the other models, having
the lowest misclassification rate of 0.363. The selected model has a sensitivity of 49.7% and a specificity
of 75.1% in the validation data.

INTRODUCTION
Data related activities in healthcare management consist of data collection, data sharing, and data
analytics. Of those three parts, data collection and data sharing have been successfully achieved by
implementing various technologies suitable for collecting and sharing health care records. One of the
prominent dimensions in health care is to design services and manage resources by understanding
patient conditions. Such situations may include not only chronic conditions a patient might be suffering
from, but also external factors such as mode of admission into the health care facility, length of stay in the
hospital, patient age, gender, ethnicity, and so on. Further foray into analytics has provided tremendous
advantages that assist in designing follow-up services specifically catering to chronic illnesses, thereby
reducing costs significantly and improving health care services.
Several measures have been taken by the Centers for Medicare & Medicaid Services (CMS) to reduce
hospital readmission rates and improve the quality of healthcare in the United States. According to CMS,
one of the major issues faced by our healthcare system is avoidable readmissions which cost the
Medicare approximately $15 billion per year. On average, Medicare spends double the amount for an
episode with one readmission. In order to handle this situation, hospitals are required to publicly report
their re-admission rates and those with high readmissions have to pay a re-imbursement penalty as well.
Conditions with highest readmission rates were acute myocardial infraction, heart failure, pneumonia and
diabetes. Everybody is different and so are their medical needs. A standard umbrella prescription of diet,
regimen, and medication may not be suitable for patients with such different conditions. A custom
approach for each medical condition is required to identify patients who have higher likelihood of being
readmitted [11].
Diabetes is a chronic condition which serves as a precursor for other medical conditions and could trigger
other health concerns in future. It has been estimated that about 366 million people worldwide have
diabetes, and this number is likely to escalate to 552 million by the year 2030 [2].
Around 25.8 million people in the U.S., which is about 8.3% of the population at present, have diabetes. If
the present trend continues, one in every three U.S. adults is likely to have diabetes by the year 2050 [1].
From a monetary consideration, in 2012 alone diabetes has cost the U.S. approximately $245 billion
dollars [1] and unchecked diabetes resulted in 23,700 readmissions and cost $251 million [12]. SAS

1
Enterprise Miner 13.1 is used in this paper to identify patients among various surgical groups who display
a higher likelihood of readmission.

LITERATURE REVIEW
Health care organizations continue to develop optimal solutions in chronic care models. Examination of
scholarly articles has revealed that out of 39 studies, 32 interventions based on chronic care models have
resulted in improvement of at least one outcome measure according to Thomas Bodenheimer, MD,
Edward H. Wagner, MD, MPH and Kevin Grumbach, MD [6]. From a financial perspective, studies
suggest that out of 27 cases, 18 have confirmed that using chronic care models has reduced health care
costs and led to a decrease in the use of health care services [6].
According to the study carried out by Allen M. Glasgow, Jill Weissberg-Benchell, W. Douglas Tynan,
Sandra F. Epstein, Chris Driscoll, Jane Turek, EvBeliveau in “Readmissions of Children with Diabetes
Mellitus to a Children's Hospital”, it was observed that readmission of diabetic children increased due to
missed insulin injections [9].
Another point of view is presented by BeataStrack, Jonathan P. DeShazo, Chris Gennings, Juan L. Olmo,
Sebastian Ventura, Krzysztof J. Cios and John N. Clore in “Impact of HbA1c measurement on Hospital
Readmission Rates: Analysis of 70,000 Clinical Database Patient Records” where the authors used the
multivariate regression method to conclude that greater attention to HbA1c determination may improve
outcomes and reduce costs [8].
Using such chronic illness management models to develop personalized interventions in hospitals has
yielded better patient outcomes, decreased length of stay, and lessened rates of re-admission in a 30 day
period. Jill Koproski, RN, CDE, ZoraydaPretto, MD and Leonid Poretsky, MD, conducted a random
feasibility study on the effects of an intervention in hospitalized diabetics by a medical research team and
found that it resulted in decreased length of stay in the hospital as well as an overall improvement in the
glycemic levels [7].

DATA
This study uses the data obtained from the Center for Machine Learning and Intelligent Systems at
University of California, Irvine [4]. It contains clinical records from over 100,000 individual encounters
corresponding to over 60,000 distinct patients. The data was collected from over 74 million encounters
related to 17 million patients [8]. It was collected over a period of 10 years, from 1999 to 2008, and
contains several attributes which correspond to the times of admission and discharge of diabetic patients.
These records contain information about various laboratory tests and procedures, diagnosis, and
medications that were administered in the duration of the hospital stay. Following are the variables
selected for modeling:

Variable Name Measurement Level Range of Values

A1Cresult Nominal >7, >8, none, norm
admission_source_id Nominal 1 to 25

admission_type_id2 Nominal 1 to 4
admitted_30days Binary 0 to 1
age Nominal [0-10]... [90-100]
change2 Binary 0 to 1
diabetesmed2 Binary 0 to 1
Diagnosis_1 Nominal 3 to 999
Diagnosis_2 Nominal 5 to 999

2
 Diagnosis_3 Nominal 3 to 999
discharge_disposition_id Nominal 1 to 28
discharge_disposition_id_2 Nominal 1 to 17
encounter_id Interval 12522 to 443867222

gender Char Male, Female, Other

medical_specialty Nominal 1 to 23
num_lab_procedures Interval 1 to 132
num_medications Interval 1 to 81
num_procedures Interval 0 to 6
number_diagnoses Interval 1 to 16
number_emergency Interval 0 to 76
number_inpatient Interval 0 to 21
number_outpatient Interval 0 to 42
patient_nbr Interval ID
Race Nominal Asian, Caucasian…
Re_admitted Binary 0 to 1
Readmitted Nominal No, <30, >30

time_in_hospital Interval 1 to 14
Weight Nominal [150-170]…[170-200]
weight_new Interval 165, 175…
24 features for medications Nominal/Binary Metformin, Glipizide…… etc.

Table 1: List of variable names, measurement level, description and their range of values
Health Insurance Portability and Accountability Act (HIPPA) requires that any personal information which
would identify a patient must be kept confidential. In this research it was ensured that the dataset was
HIPPA compliant. In addition, the research conducted did not require any form of consent from the
Helsinki Declaration, enabling its exemption from the VCU Institutional Review Board review [8].
The data set “Diabetes 130-US hospitals for years 1999-2008” incorporated the use of International
Statistical Classification of Diseases and related health Ppoblems, also known as ICD [10]. ICD is
maintained by World Health Organization (WHO) and serves as a “standard diagnostic tool for
epidemiology, health management and clinical purposes” [10]. ICD serves as a data dictionary for various
kinds of health conditions. Furthermore, to explain the health condition classification system, it provides
detailed explanation of the conditions, for instance signs, symptoms, and so on [10]. The dataset used for
this study incorporated the majority of the ICD systems ranging from ICD-9001 to ICD-9999 [10].

DATA PREPARATION:
The original dataset had 55 attributes and 101,731 observations. In order to prepare the dataset for
modeling, the data was subjected to intense cleaning procedures. The final dataset created has 91,841
observations and 53 attributes as displayed below.

3
Figure 1 Variable Summary
The nominal variables such as diagnosis, race, medical specialty, and discharge disposition had missing
values. These missing values need to be handled prior to being fed into the model as they may introduce
instability and bias in models such as regression and support vector machines. In order to avoid imputing
missing values in primary, secondary and tertiary diagnosis, the observations that had missing values in
them were removed from the analysis using PROC SQL. ‘Medical Specialty’ variable, containing
approximately 50% missing values, pertains to the medical specialty of the physician who attended the
patient and prescribed the initial diagnosis. ‘Weight’ variable has 97% missing values. After careful
consideration, weight, payer code, and medical specialty variables have been rejected in model building.

Figure 2 Interval variable summary statistics

Figure 3 Class Variable Summary Statistics

Some variables, for example weight, race, and encounter ID contained anomalies such as questions
marks and other special characters. Observations with invalid gender values were deleted. The variables

4
discharge disposition and race had few percentages of missing values and they were imputed using the
Tree imputation method in the Impute node. The nominal variable ‘discharge disposition’ had 29 levels
such as null, unknown, and other etc. These levels were grouped in a more logical way and reduced to 6
levels using SAS code. Similarly, the variables admission type and admission source had 8 and 26 levels.
As many levels were similar to each other with subtle differences, they were categorized into 4 and 7
levels respectively. In addition, 11 categorical variables were converted into binary variables for ease of
analysis. The data was checked for outliers via box and whisker plots. The observations that were beyond
the range of ±3 standard deviation were deemed as outliers and were accordingly removed using SAS
code. The target variable of readmission had three levels of ’<30’ (patient readmitted within 30 days),
‘>30’ (patient was readmitted after 30 days) and ’no’ (patient was not readmitted). In this paper, the target
variable has been re-coded as a ’1’ (Patient is readmitted) and ‘0’ (patient is not readmitted) as a binary
variable.
The interval variables ‘number emergency (number of times a patient was admitted via emergency)’ and
‘number outpatient (number of times the patient was an outpatient before being admitted)’ had high
values of skewness and kurtosis. In order to normalize these variables, log transformation was used to
reduce the skewness and kurtosis values of these variables.

MODEL BUILDING
The data was partitioned into train (60%) and validation (40%) datasets to provide an honest model
assessment.. As the percent of target variable in each outcome is balanced in this data, there is no need
to set prior probabilities.

Display 1 Initial process flow

SAS Enterprise Miner 13.1 provides many new features and nodes. The newer sets of models such as
high performance nodes are available in this version and used in the analysis. The primary purpose of
this paper is to predict whether or not a patient will be re-admitted and to identify the key factors serving
the objective. As the target is a binary variable, the selection criterion used was the validation
misclassification rate. For variable selection purposes, the variables selected from different nodes such
as the HP variable selection node with variations (LARS, LASSO, and Adaptive LASSO), LARS node, the
HP GLM node and the Decision Tree node were used to identify the significant predictors. Among all
these variable selection nodes, the variables selected by the LARS model were most instrumental in
decreasing the validation misclassification rate.
In order to predict the patient outcome measures, various models were built such as the Decision Tree
with variations in the splitting rule target criteria (default, Entropy, Gini, HP Tree), the Regression models
with variations in variable selection criteria (default, decision tree, stepwise, HP Tree), the Gradient
Boosting with default properties, the Rule Induction node where the binary model used is Tree and the
cleanup model used was ‘Neural’, the MBR model using the RD-Tree method with 16 neighbors, and the
Support Vector Machine node with the Linear Kernel Function. The model comparison node has been
used to select the model with the least validation misclassification rate. Among all the models, the LARS
model (variable selection) used in conjunction with the HP Support Vector Machine had the lowest
validation misclassification rate of 0.363.

5
Display 2 Process Flow Diagram

Output 1 Fit Statistics of Model Comparison Node

EXPLAINING THE BEST MODEL- HP SVM WITH VARIABLE SELECTOR (LARS)

LARS (Least Angle Regression) is used to identify the important predictors for model building. It is a
recently developed specialized regression technique that is used for variable selection as well as model
fitting and prediction. LARS can handle both numeric and categorical input variables. Using LARS,
problems such as violation of hypothesis assumptions and selection bias may be overcome by using
information criteria such as AIC and SBC.

6
Output 2 Parameter Estimate (Absolute Value) LARS
Shown above is the plot of absolute values of the parameter estimates from the LARS model. The
variable ‘Number_Inpatient’ in the Parameter Estimate graph has the highest estimate of 0.158. The
important predictors in order of decreasing importance are as follows:
 The number of times a patient has been an inpatient previously has a positive effect on readmission
followed by the number of times he/she was an outpatient.
 The number of times a patient was admitted via the emergency room has the next highest positive
impact with a likelihood of 4%.
 Diabetic patients suffering from a primary diagnosis of heart failure have a higher chance of
readmission.
 Number of diagnoses entered into the system corresponding to each patient is one of the important
predictors.
 Diabetic patients who were transferred from another hospital, home health agency, ambulatory
surgery center, critical access hospitals, or due to court/law enforcement have reduced chances of
being readmitted.
 Patients who were transferred from a skilled nursing facility (SNF) were discovered to have increased
chances of readmission.
 In addition, patients suffering from a secondary diagnosis of hypertensive chronic kidney disease
demonstrated increased readmission rates.
 Patients who were discharged/ transferred to home with home health service had a positive effect,
thus indicating a higher chance of coming back to the hospital.
 Diabetic women with complicated pregnancy or childbirth have displayed a reduced likelihood of
being readmitted.
 Patients suffering from a primary diagnosis of chronic bronchitis are more likely to be readmitted; it
was also noticed that patients with diabetes mellitus had higher chances of being readmitted.
However, the LARS model is used as a variable selector and how these selected effects are weighed in
to the final SVM model is unknown.

7
Support Vector Machine (SVM) algorithm has been selected as the winner model by the model
comparison node since it has the lowest validation misclassification rate of 0.363. It is a supervised
machine learning method which is used to solve binary classification problems. Generally, the data in the
real world is not linearly separable. SVM maps the data into a higher dimensional space (infinite) which
helps in separating the data points more effectively. Inside SVM, for a data-point having n-dimensions (n
input variables), a (n-1) hyperplane (a linear combination of the input variables) is generated. This (n-1)
dimensional hyperplane is generated using a kernel function which is used to separate the data points in
a yes/no, 0/1 or +/- situation. The kernel function used here is the Linear Kernel function
The SVM node takes the training data, selects the weight and bias in such a way that the hyperplane
separates the training data points. Once the plane is known, if the plane is on the positive direction, it is
termed as positive data; if the plane is on the negative direction, it is termed as negative data. The
purpose of SVM is to find the best hyperplane, which maximizes the margin or the boundary vectors
(vectors define the boundary of the classes). The best hyperplane provides maximum safety in case the
predicted value is incorrect. The optimization problem is to maximize the distance or minimize the width
(w) of the hyperplane. Under circumstances where the data is not completely separable, the SVM model
imposes a penalty. The idea is to allow penalty for classifications which were wrongly predicted. The
following variables were fed into the model: 7 interval variables and 16 class variables (2,042 class
variable levels). The number of effects selected by the model is 23.

Output 3 SVM Fit Statistics

The accuracy obtained by the model is 63.38% for the validation data. With the help of the model and
sorting the observations based on probabilities, it can be stated that patients in the top decile have 61.6%
higher likelihood of being readmitted.

8
Output 4 Cumulative Lift SVM

Output 5 Fit statistics SVM

CONCLUSION
The hospital data of in-patients having diabetes as an existing condition in conjunction with other medical
illnesses was analyzed to build a predictive model to identify patients who had a higher likelihood of being
readmitted. Some of the key factors that drove readmission are the number of times a patient was
formerly admitted both as an inpatient and outpatient, primary diagnosis, mode of admission, conditions
like heart failure or hypertensive chronic kidney disease and so on. Using the model, it can be stated that
patients in the top decile have 61.6% higher likelihood of being readmitted.

DISCLAIMER
The authors do not possess any medical expertise. The results are based purely on the data and are not
intended to serve as a medical advice.

REFERENCES
1. http://mitsloanexperts.mit.edu/using-analytics-to-manage-diabetes/
2. http://www.sciencedirect.com/science/article/pii/S0956566313004260
3. http://www.hindawi.com/journals/
4. https://archive.ics.uci.edu/ml/datasets/Diabetes+130-US+hospitals+for+years+1999-2008
5. http://archive.ics.uci.edu/ml/

9
 6. http://jama.jamanetwork.com/article.aspx?articleid=195407
7. http://care.diabetesjournals.org/content/20/10/1553.short
8. http://www.hindawi.com/journals/bmri/2014/781670/
9. http://pediatrics.aappublications.org/content/88/1/98.short
10. http://en.wikipedia.org/wiki/International_Statistical_Classification_of_Diseases_and_Related_He
alth_Problems
11. https://www.healthcatalyst.com/healthcare-data-warehouse-hospital-readmissions-reduction
12. https://www.fiercehealthfinance.com/story/readmission-lead-413b-additional-hospital-costs/2014-
04-20

CONTACT INFORMATION
Your comments and questions are valued and encouraged. Contact the authors at:
Dr. Goutam Chakraborty
Oklahoma State University
Stillwater, OK, 74078
goutam.chakraborty@ okstate.edu

Dr. Goutam Chakraborty is Ralph A. and Peggy A. Brenneman professor of marketing and founder of
SAS® and OSU data mining certificate and SAS® and OSU marketing analytics certificate at Oklahoma
State University. He has published in many journals such as Journal of Interactive Marketing, Journal of
Advertising Research, Journal of Advertising, Journal of Business Research, etc. He has chaired the
national conference for direct marketing educators for 2004 and 2005 and co-chaired 2007 data mining
conference. He has over 25 years of experience in using SAS® for data analysis. He is also a Business
Knowledge Series instructor for SAS®.

Hephzibah Munnangi
Oklahoma State University
Stillwater, OK- 74075
Work Phone: 405-385-4992
Email: [email protected]

Hephzibah Munnangi is a Master’s graduate in Management Information Systems from Oklahoma State
University. She has a bachelor’s degree in electronics and communication engineering. She has two
years’ experience in using SAS® tools for Database Marketing and Predictive Modeling. She is a Base
SAS® 9 certified professional, SAS Certified Statistical Business Analyst Using SAS 9: Regression and
Modeling Credential, JMP Software data exploration certified and holds the SAS and OSU Data Mining
certification. She is currently working as an Analyst at Ipsos MMA.

SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of
SAS Institute Inc. in the USA and other countries. ® indicates USA registration.
Other brand and product names are trademarks of their respective companies.

10