Intro: Poor neurologic outcome in extremely low birth weight(ELBW; <1000g) infants is common and its determinants have been difficult to identify. Neural networks are nonparametric pattern recognition techniques useful in detecting complex non-linear relationships.

Aim: To identify the major determinants of major handicap (Any of: Cerebral Palsy /Blindness /Deafness /Mental Retardation /Hydrocephalus / severe Failure To Thrive) and low Bayley Scores (Corrected MDI and PDI) at 18 months of age using regression techniques and neural networks.

Method: A composite database was created by combining the NICU database (21 variables of antenatal and postnatal events) with the newborn followup database for the years 1990-94. ELBW infants with followup to 18 months (n=218) were randomly divided 2:1 into a training set (n=144) and a test set (n=74). The training set was used to develop multiple logistic regression (handicap), linear regression (MDI, PDI), and neural network models(QNet v2; four-layer back-propagation) which were used to predict the outcome in the test set.

Results: Results obtained by regression techniques and neural network correlated strongly with each other (r=0.8, p<0.001). However, both regression techniques and neural networks had low sensitivity (for handicap) and r values (for MDI/PDI) (TABLE). Major determinants of increased risk (p<0.05) were: Handicap: IVH Grade, NEC ≥ Stage II, Black Race, No Chorioamnionitis

Table 1 No caption available.
Full size table

MDI: IVH Grade, BPD, Plurality, No Chorioamnionitis, Maternal education (lower grade levels)

PDI: IVH Grade, PVL, BPD, No Chorioamnionitis, Maternal education(lower grade levels)

Conclusions: The majority of the variance in neurologic outcome cannot be explained by risk factor-based statistical approaches, including non-linear neural networks. More wide-ranging and detailed data collection and analysis may be necessary to more accurately predict outcome.