Background: Clinical detection of intrapartum hypoxic-ischemic encephalopathy (HIE) from electronic fetal monitoring signals is limited by subjective interpretations with low specificity. Automated classification techniques seek to address this. This study focuses on a fetal heart rate (FHR) signal representation for classification using scattering transform (ST). This nonlinear time-scale filtering using wavelets has attractive stability, time-invariance, and discriminative properties. The hyperparameter T, which controls ST lowpass-filter width before downsampling, determines the degree of time invariance of ST coefficients.
Aim: This study investigates the sensitivity of ST T value in a classifier for early detection of fetuses at risk of HIE during labour.
Method: Our data consisted of up to 12 hours of FHR from 418 healthy and 417 HIE neonates from KP hospitals. We applied ST with maximum wavelet scale J=11 to 20-min epochs of FHR and used the coefficients to train a 3-layer bidirectional long short-term memory to classify fetuses as either HIE or healthy. Using 10-fold cross-validation with 10 repetitions, the model with average specificity of 0.7 and highest average sensitivity for each fold was selected. We chose T by varying δJ from 1 (usual default) to J. At each epoch, we used Kruskal–Wallis' test to determine whether performance differences for T value pairs were significant.
Results: Figure 1 shows test data AUROC median, 10th and 90th percentiles over time for T=64 and T=1024. The black asterisks indicate significant differences between distributions. AUROCs for T=16, 32 and 64 were not distinguishable, but higher T generates fewer ST coefficients, so is preferable. For T=64 and T=1024 there were significant differences as early as 10 hours before delivery.
Conclusion: This study demonstrates that adjusting ST T values to achieve higher time resolution produced better prediction. In future, we will conduct more hyper-parameter tuning and extend current analysis to larger imbalanced dataset.