The broad purpose of this proposal is to create a framework for bedside decision support to predict life threatening events before they happen. The specific hypothesis is that models predicting cardiac arrest can be generated from physiologic and laboratory data obtained in the 12 hours preceding the event using logistic regression analysis (LR) and data mining techniques such as support vector machines (SVM), neural networks (NN), Bayesian networks (BN) and decision tree classification (DTC). We further hypothesize that a support vector machine technique will yield the model with the best performance.
Specific Aim 1 is to acquire and prepare data for eligible patients by merging information from physiologic, laboratory, and clinical databases and selecting data from twelve hours prior to either a cardiac arrest or the maximum severity of illness. Noise will be removed with automated methods that can be used in real time. Missing data elements will be imputed by statistical methods that are regarded as state of the art. Since the optimum time window to investigate before an arrest has not been established, and since there is no standard process of abstracting trend information, we will generate multiple candidate data sets in an effort to determine the optimum combination of parameters. Data dimensionality will be reduced by three separate feature selection methods, each of which will be used in subsequent modeling procedures.
Specific Aim 2 is to create cardiac arrest prediction models from the candidate data sets using LR, SVM, NN, BN and DTC. We will assess model performance with sensitivity, specificity, positive predictive value, negative predictive value, and area under the Receiver Operating Characteristics curve (AUROC) using 10- fold cross validation. We will then assess the ability to generalize by testing the model on unseen data. We will determine the impact of training sample size on model performance by varying the percentage of data used during the 10-fold cross validation for each modeling technique's best performing model. We will then perform a false prediction analysis to determine the etiology of the false prediction.
Specific Aim 3 is to determine which modeling process and configuration parameters performs the best, and to determine optimum timing windows for: time to analyze pre-arrest and size of feature window. The significance of this proposal is that successful prediction and early intervention could save thousands of lives annually. ? ? ?
| Kennedy, Curtis E; Aoki, Noriaki; Mariscalco, Michele et al. (2015) Using Time Series Analysis to Predict Cardiac Arrest in a PICU. Pediatr Crit Care Med 16:e332-9 |
| Kennedy, Curtis E; Turley, James P (2011) Time series analysis as input for clinical predictive modeling: modeling cardiac arrest in a pediatric ICU. Theor Biol Med Model 8:40 |
