Exploratory Data Analysis to Detect Preterm Risk Factors. Abstract and Executive Summary of Dissertation;Rept. for 1 Sep 96-31 Aug 97

摘要

The study tested whether data-mining techniques applied to a comprehensive clinical practice database would detect previously unrecognized factors predictive of preterm birth. Coded and numerical clinical data from 19,970 births were extracted from a regional perinatal electronic medical record system, resulting in raw data sets containing available variables at conception (441 input variables), 10-weeks (834), 20-weeks (1,227), and 30-weeks (1,620) gestation. After applying data-reduction techniques, deliveries were randomly assigned to training and testing sets at 75% and 25% respectively, and the training sets were used separately to create prediction models using multiple linear regression, logistic regression, and neural networks. The 48 models were then used to generate predictions for the test cases. Area under the ROC curve for the models ranged from .6468 (SE = .0101) for linear regression on raw data to .7058 (SE = .0107) for neural networks using a genetic algorithm-reduced data set of raw data and principal component scores.