Support vector regression applied to foetal weight estimation

Foetal weight (FW) estimation based on echographic measurements has paramount importance in delivery risk assessment. Our previous experiments (2000) have shown that foetal weight prediction can be achieved with lower errors with MLP and RBF neural nets than those obtained with classical linear regression models. Our best model tested with our FW data set achieved a mean relative error of 6.2%. The paper reports experiments using SV machines in an editing step to data smoothing and in a regression approximating step to predict FW from three echographic features. The averaged relative mean error in the range [2000, 4500] grams was 5.0% and the average percentage of estimated FWs whose relative error was less than 5% of the FW was 67%. These results seem to be a good contribution to the research objective, which is to know to what extent combining neural nets can improve over the 15% error of FW estimation using prediction formulas in current day clinical use.