A decision support tool for allocating hospital bed resources and determining required acuity of care
Introduction
Hospitals are faced with limited resources including beds to hold admitted patients. This resource constraint is particularly important in specialized areas of the hospital, such as intensive care (ICU) and intermediate care units, since the number of beds available is a fraction of the number of floor beds, where the intensity of care is less. Cost-effective patient management is critically dependent on accurate assessment of individual patient outcome and resource utilization [3]. Evaluating length of stay (LOS) information is a challenging task [21], but is essential for the operational success of a hospital. Intensive care resources in particular are often limited and pose scheduling problems for hospital staff and administrators [18]. Predicting LOS is difficult and is often only done retrospectively. There are different components of a patient's LOS. Lengths of stay may be evaluated for ICUs, step down units, and floor units individually, and may also be evaluated as an aggregation of all areas to provide a total LOS. Existing research in this area looks at predicting or analyzing the LOS for patients in a specific hospital area (e.g., ICU) [3], [7], [10], [11], [18]. The research presented in this article will initially be concerned with ICU LOS, but will ultimately address predicting the LOS for a patient at all levels of care in the hospital and providing a total LOS.
A patient's LOS is correlated with that patient's injury or illness severity [9], that is patients who are more critically ill normally require greater lengths of stay within hospitals to return to a functional level. Therefore, it is reasonable to assume that a single decision support model will be able to estimate both of these values for patients with a given set of input variables. Establishing a patient's injury severity may require extensive and invasive testing of the patient.
Earlier identification of injury severity, and consequently LOS, would enable better resource planning and ultimately improve patient care by the hospital. The likely correlation between injury severity and LOS indicates that if characteristic variables of injury severity can be identified, they can also be used to model LOS. Since the relationship between the characteristic variables (e.g., age, type of injury, history of heart disease, …) and injury severity (and consequently LOS) is unknown, a nonparametric modeling technique is desired. Additionally, the modeling of injury severity or LOS appears to be a categorization problem. Prediction of the acuity of care required by a patient (or injury/illness severity) naturally separates the population of patients into several distinct categories (discussed in more detail in 3.3 Neural network hidden-layer architectures, 4 Results). Those patients that have critical injuries and must be treated in an ICU unit are members of one category of patient and patients with more moderate injuries and only requiring care in a standard floor unit are members of another category. Neural networks are an ideal tool for creating nonparametric categorization models.
The research in this article utilizes a supervised learning neural network approach to create a decision support tool that accurately predicts the LOS for two distinct patient populations: (1) pediatric trauma patients and (2) patients with acute pancreatitis. Patients with acute pancreatitis are initially treated in a hospital's emergency room (ER), but are not considered trauma patients. Additionally, the neural network models will define the acuity of care needed (or injury severity) for the pediatric trauma patients. Severity measures were not available for the acute pancreatitis patients, however, domain experts have communicated with the research team that illness severity is normally categorized as a direct function of the LOS (i.e., S=f(LOS)+ε).
Data for the pediatric trauma patients are taken from the National Pediatric Trauma Registry, which currently contains over 8000 records of patients that have been treated at a participating hospital's emergency room, while data for the patients with acute pancreatitis are obtained from patient records at a single hospital. The research will attempt to confirm the efficacy of using neural networks with information that is available within the first 10 min (without the use of invasive tests) of the patient's arrival at an emergency room to accurately predict a patient's LOS and the severity of the patient's injuries.
Section snippets
Background and significance
Neural network applications in medicine have been primarily limited to laboratory applications (e.g., test evaluation) or medical imaging applications [6], [14]. Buchman et al. [3] reports that neural networks in patient assessment domains typically outperform standard statistical modeling techniques. Recently, a few applications of neural networks, described in the ensuing paragraphs, have been made in the area of patient planning and medical resource allocation for patient care.
Previous
Neural network models for LOS prediction
In this section, the methodology used to design and implement the respective neural network models is discussed. The neural network design methodology follows the guidelines for neural network implementation made by Walczak and Cerpa [20].
Results
Separate neural network models are developed and implemented for each of the two medical domain problems, as described in the previous section. Recall that the pediatric trauma model covers numerous types of trauma categorizing the patient into one of five categories (four of which are presented due to the lack of data in the fifth category) and the acute pancreatitis model places patients into three unique categories. Each model is presented in turn.
Neural network evaluation of LOS heuristics
The results presented in the previous section lend strong support for validation of the research goal: that a patient's LOS and severity of injury or illness may be accurately predicted by a neural network model using only information that is available within 10 min of a patient's arrival at a hospital's emergency room. Additionally, a very interesting corollary result is obtained from the acute pancreatitis neural network decision support models. In addition to removing input variables that
Summary
The research reported in this article has shown that neural network systems are capable of providing significant planning and patient care information with a relative paucity of diagnostic knowledge. All knowledge used as input for the neural networks implemented and tested in the presented research is available within the first 10 min of a patient's arrival at the hospital for the pediatric trauma patients and within the first 2 h after arrival for the acute pancreatitis patients (the optimal
Steven Walczak is a faculty member in the College of Business and Administration at the University of Colorado at Denver. He received his PhD from the University of Florida and his MS and BS from the Johns Hopkins University and Pennsylvania State University, respectively. Dr. Walczak is an active member of the Association for Information Systems, the Decision Sciences Institute, and belongs to the American Association for Artificial Intelligence (AAAI). His current research interests are in
References (24)
- et al.
Efficacy of modified backpropagation and optimization methods on a real-world medical problem
Neural Networks
(1995) - et al.
Extrapolation and interpolation in neural network classifiers
IEEE Control Systems
(1992) - et al.
A comparison of statistical and connectionist models for the prediction of chronicity in a surgical intensive care unit
Critical Care Medicine
(1994) - et al.
A fuzzy ARTMAP nonparametric probability estimator for nonstationary pattern recognition problems
IEEE Transactions on Neural Networks
(1995) - et al.
Prediction of rib fracture injury outcome by an artificial neural network
Journal of Trauma: Injury, Infection, and Critical Care
(1995) - et al.
Artificial neural networks in pathology and medical laboratories
Lancet
(November 4, 1995) - et al.
Simulated biologic intelligence used to predict length of stay and survival of burns
Journal of Burn Care and Rehabilitation
(1996) - et al.
Multilayer feedforward networks are universal approximators
Neural Networks
(1989) - et al.
Does severity explain differences in hospital length of stay for pneumonia patients?
Journal of Health Services Research and Policy
(1996) - et al.
Prediction of trauma mortality using a neural network
American Surgeon
(1997)
Predicting length of stay for psychiatric diagnosis-related groups using neural networks
Journal of the American Medical Informatics Association
Design and Development of Expert Systems and Neural Networks
Cited by (0)
Steven Walczak is a faculty member in the College of Business and Administration at the University of Colorado at Denver. He received his PhD from the University of Florida and his MS and BS from the Johns Hopkins University and Pennsylvania State University, respectively. Dr. Walczak is an active member of the Association for Information Systems, the Decision Sciences Institute, and belongs to the American Association for Artificial Intelligence (AAAI). His current research interests are in the areas of applied intelligent decision support systems, including neural networks and expert systems, in the domains of finance and medicine. He is previously published in Journal of Management Information Systems; Decision Support Systems; IEEE Transactions on Systems, Man, and Cybernetics; Expert Systems with Applications, and many other journals.
Walter Pofahl is an Assistant Professor in the Department of Surgery of the Brody School of Medicine at East Carolina University. He received his medical degree from West Virginia University. His postgraduate surgical training was at the University of Kentucky Medical Center, Lexington, KY. His areas of clinical interest and expertise are General and Laparoscopic Surgery. His research interests include outcome assessment in acute pancreatitis. He has presented at numerous national meetings. His work has been published in the peer-reviewed medical literature.
Ronald Scorpio is a physician with the Spartanburg Regional Healthcare System. His research interests are in pediatric trauma. He has recently been published in Archives of Pediatric and Adolescent Medicine, Pediatric Critical Care Medicine, and various other medical journals.