Rdf. Keith et al., SUITABILITY OF ARTIFICIAL NEURAL NETWORKS FOR FEATURE-EXTRACTION FROMCARDIOTOCOGRAM DURING LABOR, Medical & biological engineering & computing, 32(4), 1994, pp. 190000051-190000057
Fetal condition during labour is inferred from a continuous display of
fetal heart rate and uterine contractions called the cardiotocogram (
CTG). The CTG requires a considerable expertise for correct interpreta
tion, which is not always available. We are developing an intelligent
system to support clinical decision-making during labour. The system's
performance depends on its ability to classify features from the CTG
similarly to experts. Artificial neural networks (NNs) can be taught b
y experts for such tasks, and so may be particularly suitable. We foun
d NNs suitable for feature extraction when the problem was reduced to
small well defined tasks, and numerical algorithms were used to pre-pr
ocess the raw data before application to the NNs. A NN with optimised
dimensions was used in this way to classify the magnitude of decelerat
ions, a feature clinicians find particularly difficult. The NN was com
pared with the algorithm used in a commercial antenatal monitor and s
ix reviewers which included two CTG experts. The experts were consiste
nt (89.7% and 97.0%) and agreed well with each other (81.0%), whereas
the non-experts were less consistent and agreed less well The NN agree
d well with the experts (75.0% and 81.9%) but the algorithm agreed poo
rly (56.5% and 68.9%). It was found that the algorithm's performance c
ould be improved (72.1% and 76.7%) when modified to use additional inf
ormation. Our earlier attempts to fully classify the raw CTG using a s
ingle NN were unsuccessful because of the large number of data pattern
s. A simplified approach to classify the magnitude and timing of decel
erations was also unsuitable when contraction data was Of poor quality
or absent. We have adopted a hybrid approach for our intelligent syst
em for reliability and improved performance. CTG features are extracte
d and classified by combined numerical algorithms and small NNs. These
features, together with patient information, are then processed by an
expert system which allows interaction with clinicians.