Automated evaluation of fetal cardiotocograms using neural network

Abstract

Cardiotocograms (CTG) monitor fetal heart rate and uterine contractions and are routinely used as a diagnostic tool by obstetricians to determine fetal state. Evaluation of CTG traces require an extensive visual interpretation subject to extensive inter and intra observer variability. A neural network implementation would automatically evaluate fetal state via CTG parametric inputs in a short period of time and reduce resource costs associated with electronic fetal monitoring. The optimal neural network designed included a topology of four hidden layers with 200 neurons per layer, a scaled conjugate gradient back propagation method, and a threshold of 1.915. The optimized network performed with an absolute accuracy of 84.26%, a positive error of 10.18%, and a negative error of 5.56%. Conceptually the network performs such that only 15.74% of patients are misclassified, and of that percentage, only 5.56% of those are incorrect decisions could potentially harm the patients. Overall, substantial costs in resources and medical expertise are obviated by preventing the unnecessary investigation of 69.9% of all patients. Despite its scholastic merit, the neural network may not be suitable for medical incorporation due to strict FDA regulations and political and public scrutiny over endangerment of 5.56% of all fetal patients. Use of a developmental parameter could potentially reduce error and make such an interactive decision support tool feasible and should be investigated.