Table 3 The comparison between our study and previous studies.
From: Automated interpretation of cardiotocography using deep learning in a nationwide multicenter study
Study ID | Pre-processing method | Type of features extracted | Part of CTG used | Clinicians as author(s) | Hold-out-validation | Cross validation method | ML classifier(s) used | Model interpretability | Performance measure(s) | Oversampling |
|---|---|---|---|---|---|---|---|---|---|---|
Our study | Excluded records with missing data (e.g., maternal age, Apgar scores) or incomplete twin signals (> 1-min discontinuity) | From 0.5 Hz time-point signals, 7 statistical features (max, min, median, mean, number of peaks, variance, standard deviation) were extracted for the final model | FHR and UC | Both | Yes | None | 1D-SEResNet50 | No | SEResNet50: | No |
Specificity = 81% | ||||||||||
NPV = 94.7% | ||||||||||
AUPRC = 62.5% | ||||||||||
AUROC = 88% | ||||||||||
Linear interpolation | FHR FIGO features and UC | FHR and UC | Non-clinician | Yes | k-fold | LG | Yes | LG: | No | |
AUROC = 74% | ||||||||||
Did not specify | Morphological | FHR and UC | Non-clinician | No | k-fold | MLP, bagging, RF ands SVM | Partially | RF | Yes | |
Sensitivity = 96.4% | ||||||||||
Specificity = 98.4% Accuracy = 96.7% | ||||||||||
Precision = 96.8% | ||||||||||
Remove spikes, interpolate, and segment into 20 min | Wavelet packet decomposition image | FHR | Non-clinician | No | k-fold | 2DCNN | No | CNN: | No | |
Accuracy = 95.24% Sensitivity = 90.4% Specificity = 100% | ||||||||||
Smoothing | Morphological and statistical | FHR and UC | Non-clinician | No | k-fold | NN, RF, clustering and SVM | Partially | Ensemble combination- NN, RF, k-means and SVM: Accuracy = 92.30% | No | |
Processing outliers and removing spike using moving average | Image | FHR and UC | Non-clinician | Yes | Did not specify | 1D-CNN and bidirectional Gate | No | Accuracy = 95.15% | No | |
Sensitivity = 96.20% Specificity = 94.09%, Precision = 94.21% | ||||||||||
Recurrent Unit (BiGRU) | F measure = 95.20% | |||||||||
AUROC = 99.29% | ||||||||||
Outlier detection and linear interpolation | Linear and nonlinear, extract feature using CNN & LSTM | FHR | Non-clinician | No | k-fold | SVM and CNN-BiLSTM | Partially | SVM: | No | |
Sensitivity = 56.97% Specificity = 73.35% | ||||||||||
QI = 63.91% | ||||||||||
Did not specify | Image based and text | FHR | Non-clinician | Yes | Stratified k-fold | CNN | No | MMIF-1 (ViT-B/16): Accuracy = 96.3% | No | |
F measure = 96.3% | ||||||||||
AUROC = 96.2% | ||||||||||
Did not specify | Image | FHR | Non-clinician | Yes | k-fold | KNN, NB, SVM, DT, RF, ADABOOST, XGBOOST | No | XGBOOST: | No | |
Accuracy = 96.3% | ||||||||||
Precision = 95.4% | ||||||||||
Recall = 97.3% | ||||||||||
F measure = 96.4% | ||||||||||
AUROC = 95.9% | ||||||||||
Lagrange interpolation | Image | FHR | Non-clinician | No | Did not specify | Double Trend Accumulation Former CNN | No | Accuracy = 90.6% | No | |
Lagrange interpolation | Curve classification | FHR | Non-clinician | Yes | k-fold | Trend-Guided Long CNN | No | Accuracy = 89.80% | No |
