Analyzing Fetal Heart Rate Patterns via Latent Representations with Variational Recurrent Neural Networks (VRNNs)

Mahdi Shamsi1, Aditi Lahiri2, Yvonne W Wu3, Lawrence David Gerstley2, Michael William Kuzniewicz4, Marie-Coralie Cornet3, Emily Hamilton1, Philip Warrick5, Robert E Kearney1
1McGill University, 2Kaiser Permanente Division of Research, 3University of California San Francisco, 4Kaiser Permanente, 5PeriGen Canada, McGill University

Abstract

Latent representations in machine learning, play an important role in extracting and interpreting complex, high-dimensional data. In the context of biomedical signals, these latent representations are invaluable for downstream tasks. They serve as a compact, yet comprehensive encapsulation of the original data, capturing essential features and underlying patterns that may not be immediately apparent from the raw signal. Building on the foundation of Variational Autoencoders, Variational Recurrent Neural Networks (VRNNs) introduce a dynamic advancement, particularly suited for time-series data such as fetal heart rate (FHR) signals. VRNNs extend the capabilities of VAEs by incorporating the temporal dependencies inherent in biomedical signals, providing a more comprehensive encoding of the evolving patterns in the FHR signal. This study provides an analysis of the latent representation of FHR signal using scattering transform and VRNN model with respect to FHR signal features and events such as acceleration, deceleration and baseline. A data-set from Kaiser Permanente Northern California hospitals was assembled from 14,372 ten-minute epochs of FHR records from 1,012 singleton vaginal births with healthy outcome. Outcome was primarily determined from blood gases sampled from the umbilical cord. First, the scattering transform was applied with a maximum wavelet scale of 11 to transform the raw FHR signals. The first 11 scattering coefficients were then used as input for the VRNN model, which incorporates a four-layer Gated Recurrent Unit for its recurrent block. The two critical hyperparameters of the model which control the model capacity are the size of hidden states of the recurrent block and the dimension of the latent variable which were set to 66 and 9, respectively. The analysis results demonstrate that the features of the latent representation from the VRNN model are significantly correlated with the FHR event features, confirming that the latent representation effectively captures critical information from the FHR signal.