Artificial neural network for predicting intracranial haemorrhage in preterm neonates

Intraventricular haemorrhage (IVH) incidence is used to assess peri-/neonatal therapy, and to make intra- and inter-hospital quality assessments. Unbiased assessment is complicated by the amount of confounding factors. Is an artificial neural network (ANN) able to early and accurately forecast the occurrence of severe IVH in an individual patient? Is it superior to classic multiple logistic regression? We conducted an observational study on pre-existing routine data. Admission data were available from 890 preterm neonates (gestational age < 32 weeks, birthweight < 1500 g). Patients were randomly assigned to either a training, or a validation set (50%/50%). Using the training set data an ANN was trained. A second predictive model was developed by stepwise multiple logistic regression analysis. Using the validation set input data both models delivered estimates of the probability for severe IVH to occur in each individual patient. Receiver operating characteristic (ROC) curves were used to compare prognostic performance. The optimal ANN processed 13 input variables, whereas stepwise logistic regression analysis only identified five independent predictor variables. The area under the ROC curve was 0.935 for the ANN and 0.884 for the logistic regression model (p = 0.001). Adjusted for 95%, 90%, 85%, 80% and 75% specificity, the sensitivity of the ANN was significantly superior to that of the logistic regression model. Due to its ability to give an accurate prognosis based solely on admission data, a trained ANN qualifies as a tool for local quality control.