Machine Learning Model for Classifying the Results of Fetal Cardiotocography Conducted in High-Risk Pregnancies

Fig. 1. Study design. Datasets were gathered and read by a team of obstetricians. Readings were recorded to evaluate 1456 fetal cardiotocography (qualified fetal cardiotocography, left side of the flowchart), while the other fetal cardiotocography results without the reading were used to evaluate the model in clinical situations (clinical validation dataset, right side of the flowchart). For the qualified fetal cardiotocography dataset, the classification model was trained to find abnormal data. Clinical validation datasets were created to represent three clinical situations according to the time window selected from the fetal cardiotocography results. NST, non stress test; CTU-UHB, Czech Technical University-University Hospital in Brno; SQI, signal quality index; IUGR, intrauterine growth restriction.

Fig. 2. Classification model structure. As an input of the classification model, 2-channel waveform data were created based on the fetal heart rate (FHR) wave and uterine contraction wave (UC). Randomly initialised convolutional kernels were applied to the input and transformed to two features that were used to classify the fetus status by lightGBM classification machine learning model.

Fig. 3. Data pre-processing flow for transforming images to CSV format. Data pre-processing consisted of three image process techniques. Waveform pixels were extracted by using RGB channel differences. Then, the edges of the graph were identified through the Hough transform algorithm results. The value of the waveform pixels was calculated using the relative position of the pixel compared to the edge. Finally, data were saved to csv format. FHR, fetal heart rate; RGB, Red Green Blue; CSV, comma-separated values.

Fig. 4. Description of the clinical validation process. The clinical validation dataset consisted of fetal cardiotocography without reads combined with maternal demographic information (age, IUP) and postpartum neonatal status (IUGR, Apgar, Fetal distress). Based on delivery, the closest 20 minutes, random 20 minutes, and all non-overlapping time windows were selected to calculate the abnormal probability score. The calculated abnormal probability score, maternal age, and IUP were used as independent variables. Regression models were fitted on IUGR, Apgar, and fetal distress as dependent variables. NST, non stress test; IUP, intrauterine pregnancy; IUGR, intrauterine growth restriction.

Fig. 5. The receiver operating characteristic (ROC; left) and precision-recall (PR; right) curves of the internal (A and B) and external validation dataset (C and D). The results were calculated through soft-voting models (10-fold each). The value of the area under the curve is shown at the right-bottom side of the graph. AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve.