Both logistic regression models (LRs) and artificial neural networks (ANNs) have been used to identify patterns in complex data sets, one of the major goals of observational studies in medicine and biology. ANNs, self-learning, non-linear, mathematical approaches, have been put forth as an alternative to the traditional, step-wise logistic regression models.

Investigators in Muenster, Germany, retrospectively evaluated data collected from 865 preterm (<32 weeks gestation) and/or low birth weight (<1500 g) babies admitted between April 1990 and October 1996, 79 of which had an intracranial hemorrhage. After randomly dividing the data set into 2 equal groups, they created a step-wise LR model and trained an ANN using one of the data sets. Each model was then validated against the other data set, to which it had not been exposed.

The ANN devised for this study consisted of a 3-layer, perceptron-type system with 13 input nodes, each connected to 4 hidden nodes, which were all connected to 1 output node. The ANN used an adaptive gradient learning algorithm with a tangential transfer function in the hidden nodes.

The sensitivity (ability to correctly identify true cases) of the ANN was higher than the LR model at the 75%, 80%, 85%, and 90% specificity (ability to correctly identify patients without intracranial hemorrhage) levels. In addition, the area under the receiver operating characteristic curve was greater (0.935 vs. 0.884; p < 0.02) for the ANN than for the LR model. Of the 13 input variables to the ANN, all seemed to contribute significantly to the prediction except "capillary PCO2 on admission" and "gender".