Enhancing cardiotocography classification via ensemble learning and threshold optimization

Figures

• 

  Figure 1. The first dataset’s experimental results include confusion matrix outputs from (a) the RF classifier, (b) the PRF classifier, and (c) the proposed method. Each column represents instances of one class. The total ‘Normal,’ ‘Suspicious,’ and ‘Pathological’ cases are 117, 38, and 13, respectively. The RF and PRF incorrectly classified 12 out of 13 ‘Pathological’ cases. The proposed method has a 76.92% accuracy on the ‘Pathological’ case.

• 

  Figure 2. The second dataset’s experimental results include confusion matrix outputs from (a) the RF classifier, (b) the PRF classifier, and (c) the proposed method. Each column represents instances of one class. The RF and PRF incorrectly classified all 12 ‘Pathological’ cases. The proposed method has a 75% accuracy on the ‘Pathological’ case.

• 

  Figure 3. The experiments result on the third data split. Confusion matrix results from (a) the RF classifier, (b) the PRF classifier, and (c) the proposed method. The ‘Normal,’ ‘Suspicious,’ and ‘Pathological’ cases are 117, 38, and 12, respectively. In terms of the ‘Pathological’ case, The RF and PRF correctly classified one case, whereas the proposed method correctly classified 5 cases.

• 

  Figure 4. Ranked importance of all extracted features used in CTG classification. Feature bars are color-coded by their category, including time-domain (orange), frequency-domain (green), autocorrelation (purple), Jx-EMGT or EMG-inspired (blue), and clinical fetal heart rate morphological features (red). The dominance of EMG and time-domain metrics is evident among the highest-ranked inputs. Please refer to Table 15 for feature name abbreviations and their full terms.

• 

  Figure 5. SHAP feature importance for three classes (0: Normal, 1: Suspicious, 2: Pathological). Please refer to Table 15 for feature name abbreviations and their full terms.

• 

  Figure 6. Block diagram of the implementation, including the data collection and preparation and individual stages leading to the decision-making process.

• 

  Figure 7. An example of signal segmentation on recording 1074, where the original recording (in black) proceeds with denoising, missing value interpolation, and spiky artifacts detect, produces two segments.

• 

  Figure 8. Auto-threshold optimization diagram, where the inference data are from Oob (Out of bags). The optimization step does not increase computation much.

• 

  Algorithm 1. Auto-threshold optimization process