ABSTRACT
III
ABSTRACT
With the economic development and the increasing number of elderly
women, the demand for prenatal fetal monitoring is increasing. However, due
to the serious shortage of medical personnel in obstetrics, especially in poor
areas, most pregnant women cannot enjoy convenient, timely and accurate
fetal monitoring services. The intelligent interpretation of cardiotocography
(CTG) model can provide auxiliary decision-making for medical staff, help
reduce the workload of medical staff, improve the detection rate of abnormal
diseases, and reduce the incidence of abnormal fetuses and stillbirths.
However, most of the existing models based on machine learning for
interpreting of cardiotocography do not meet the clinical requirements.
Therefore, this paper studies the deep forest (DF) algorithm framework and
builds an intelligent interpretation model of cardiotocography based on deep
forest.
In this paper, the clinical data of cardiotocography are first explored and
analyzed through visualization techniques and principal component analysis
for mining data distribution characteristics; Then, the CTG data is first
preprocessed, the input features are standardized, and the output labels are
encoded with one-hot encoding; Finally, based on data exploration, the deep
forest multi-grained scanning stage is used to mine the relationship between
sample features, and then the cascade forest stage is used to integrate random
forest (RF), weighted random forest (WRF), and completely random. Forest
(CRF) and Gradient Boosting Decision Tree (GBDT) are the base classifiers
and perform deep iterations to finally get the best performance model.
The experimental results show that: in the standard CTG data set,
compared with the existing domestic and foreign traditional machine learning
models and deep neural networks, the deep forest model proposed in this study
predicts the sensitivity, F1 score and area under the receiver's operating
characteristic curve (AUC) are significant, the whole AUC value reaching
0.990, the overall accuracy is 92.64%, and the average F1 is 92.01%. In this