gms | German Medical Science

MAINZ//2011: 56. GMDS-Jahrestagung und 6. DGEpi-Jahrestagung

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V.
Deutsche Gesellschaft für Epidemiologie e. V.

26. - 29.09.2011 in Mainz

Prediction of patient admission and emergency department visit based on clinical and administrative data within an electronic medical record

Meeting Abstract

  • Bianying Song - Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hanover Medical School; Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Braunschweig, Nashville
  • William Gregg - Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Nashville
  • Yukun Chen - Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Nashville
  • Min Jiang - Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Nashville
  • Dong Wang - Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Nashville
  • Josh Peterson - Department of Biomedical Informatics, Vanderbilt University, School of Medicine, Nashville
  • Klaus-Hendrik Wolf - Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig – Institute of Technology and Hanover Medical School, Braunschweig
  • Matthias Gietzelt - Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig –– Institute of Technology and Hanover Medical School, Braunschweig
  • Reinhold Haux - Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig –– Institute of Technology and Hanover Medical School, Braunschweig
  • Michael Marschollek - Peter L. Reichertz Institute for Medical Informatics, University of Braunschweig –– Institute of Technology and Hanover Medical School, Hannover

Mainz//2011. 56. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie (gmds), 6. Jahrestagung der Deutschen Gesellschaft für Epidemiologie (DGEpi). Mainz, 26.-29.09.2011. Düsseldorf: German Medical Science GMS Publishing House; 2011. Doc11gmds498

doi: 10.3205/11gmds498, urn:nbn:de:0183-11gmds4982

Published: September 20, 2011

© 2011 Song et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by-nc-nd/3.0/deed.en). You are free: to Share – to copy, distribute and transmit the work, provided the original author and source are credited.


Outline

Text

Hospital admissions and emergency department (ED) visits generate large costs for health care. Predicting patients’ risk of admission or ED visit could improve resource planning. The objective of our research is to construct and evaluate models which predict patient admission and ED visit based on clinical and administrative data recorded within the electronic medical record within the previous year. The de-identified data used in this study was drawn from StarPanel [1] within Vanderbilt University Medical Center’s adult primary care clinic. Patients with visits in the years 2008, 2009 and 2010 were included, yielding 22969 longitudinal records. Eight variables were extracted: sex, age, #hypertension medications, #diabetes medications, admission, ED visit, appointment and primary care visit. Patient admissions and ED visits were combined into a single variable (AD&EV) using a plus function. Patients were firstly classified using a K-mean algorithm based on the different single parameters e.g.: AD&EV, appointment… Decision tree, linear regression and logistic regression were used to predict the patient AD&EV based on multiple patient parameters in the previous year. The performances of these models were evaluated by receiver operating characteristic (ROC) curves [2]. The models were constructed based on the data in 2008, 2009 and evaluated based on the data in 2009, 2010.

The decision tree model is:

AD&EV class-one: N

AD&EV class-two

| Appointment class-one: N

| Appointment class-two: N

| Appointment class-three

| | #hypertension medications class-one: N

| | #hypertension medications class-two: Y

| | #hypertension medications class-three: Y

| | #hypertension medications class-four: Y

| Appointment class-four: Y

AD&EV class-three: Y

AD&EV class-four: Y

The linear regression model is:

AD&EV(next year) = 0.001994*age + 0.022049*#hypertension medications + 0.014164*#diabetes medications + 0.059382*AD&EV + 0.003469*appointment + 0.104047

The logistic regression model is:

y = -0.0108*age - 0.1032*#hypertension medications - 0.0698*#diabetes medications - 0.3169*AD&EV - 0.0163*appointment + 0.0112*primary care visit + 1.8892

The area under the ROC curve (AUC) for the linear regression, logistic regression and decision tree models are: 0.688, 0.686 and 0.628, respectively. The linear regression model which has five input parameters shows the best performance, the decision tree model which has three input parameters shows a similar performance. The AUC values of all of the models are lower than 0.7. To achieve a better performance, the inclusion of more clinical parameters in future models appears necessary.

Acknowledgement: The research stay at Vanderbilt University was supported by the Biomedical Sciences Exchange Program and the B.Braun/IALS Fellowship Program.


References

1.
Giuse NB, Williams AM, Giuse DA. Integrating best evidence into patient care: a process facilitated by a seamless integration with informatics tools. J Med Libr Assoc. 2010;98(3):220-2.
2.
Sing T, Sander O, Beerenwinkel N, Lengauer T. ROCR: visualizing classifier performance in R. Bioinformatics. 2005;21(20):3940-1.