Abstract:
Cardiotocography is the most widely used technique to monitor and evaluate the level of fetal
distress in clinical service. Currently, Cardiotocogram signal is interpreted visually by experts
using clinical guidelines. However, CTG interpretation by this methodology has shown a high
inter-observer disagreement and low specificity, leading to a poor interpretation reproducibility.
Misinterpretation of CTG signal has a significant contribution to unnecessary caesarean
deliveries and operative vaginal delivers. Misinterpretation could also be a reason for delayed
intervention when pathologic condition happens. If the labor is not intervened on time in such
condition, irreversible damage of organ or death of fetus would occur.
Several automated model has been developed to address problem in visual interpretation of CTG.
However, some model suffers from subjective labeling criteria and identify only basic guideline
features of CTG. In addition,some models suffer from hand crafted feature extraction strategy which
leads to loss of significant physiological information and low accuracy. Furthermore, most of
models were experimented on fetal heart rate data recorded in one stage of labor rather than
conducting comprehensively for data’s of 1st and 2nd stages labor. Thus their application for fetal
heart signal recorded in other stage of labor is unknown and questionable.
In this research inter-observer agreement among local experts on visual interpretation of CTG is
also evaluated to determine the extent to which a local experts agree with each other and with the
gold standard pH test. Based on the evaluation result visual annotation of CTG signal for
automated model development is ruled out and a pH test which is a gold standard biochemical
maker for fetal distress detection is used for data labeling criteria. A pre-trained AlexNet and
ResNet models were adopted and fine-tuned to select the better performing model. The fetal heart
rate data for first and second stage of labor is trained separately and the performance of the
models was evaluated. Based on training performance ResNet 50 is selected for testing phase and
a promising accuracy result of 98.7% and 96.1 are achieved for fetal heart rate data’s of 1st and
2
nd stage of labor respectively. The developed model will have a great impact in reducing the
diagnosis errors imposed by visual interpretation and can be used as a decision support system
for physicians.
