Hierarchical Attentive Network for Gestational Age Estimation in Low-Resource Settings

Abstract

Assessing fetal development is essential to the provision of healthcare for both mothers and fetuses. In low- and middle-income countries, conditions that increase the risk of fetal growth restriction (FGR) are often more prevalent. In these regions, barriers to accessing healthcare and social services exacerbate fetal maternal health problems. One of these barriers is the lack of affordable diagnostic technologies. To address this issue, this work introduces an end-to-end algorithm applied to a low-cost, hand-held Doppler ultrasound device for estimating gestational age (GA), and by inference, FGR. The Doppler ultrasound signals used in this study were collected from 226 pregnancies (45 low birth weight at delivery) between 5 and 9 months GA by lay midwives in highland Guatemala. We designed a hierarchical deep sequence learning model with an attention mechanism to learn the normative dynamics of fetal cardiac activity in different stages of development. This resulted in a state-of-the-art GA estimation performance, with an average error of 0.79 months. This is close to the theoretical minimum for the given quantization level of one month. The model was then tested on Doppler recordings of the fetuses with low birth weight and the estimated GA was shown to be lower than the GA calculated from last menstruation. Thus, this could be interpreted as a potential sign of developmental retardation (or FGR) associated with low birth weight, and referral and intervention may be necessary.

Fig. 1:

The architecture of the proposed hierarchical attention network. It contains three main components: a) convolutional feature extractor, $G_f\left(\cdot ,{\theta }_f\right)$, b) beat encoder, $G_b\left(\cdot ,{\theta }_b\right)$, and c) window encoder, $G_w\left(\cdot ,{\theta }_w\right)$. The input Doppler signal is divided into windows of 3.75 s $\left(x_1,x_2,\dots ,x_n\right)$. The scalogram of each window is calculated before feeding the network where $ith$ window has time samples $x_{i1},\dots ,x_{iT}$ after the time-frequency feature construction.

Fig. 2:

Data collection using the developed mobile application and Doppler transducers.

Fig. 2:

Data collection using the developed mobile application and Doppler transducers.

Fig. 3:

Distribution of gestational age labels and birth weight in the data set.

Fig. 4:

Architecture of the beat encoder $\left(G_b\left(\cdot ,{\theta }_b\right)\right)$ in the $\text{CNN}+\text{GRU}+{\text{Att}}_{\text{time}}$ experiment. In this structure Max Pooling was applied on the frequency dimension and attention mechanism was applied on the time dimension.

Fig. 5:

Five-fold cross-validated results using random batch generator (RBG) and strategies to deal with the imbalanced data using balanced batch generator (BBG) and BBG with balanced loss function (BBG+BLF).

Fig. 6:

Five-fold cross-validated results using different structures in the beat encoder $\left(G_b\left(.,{\theta }_b\right)\right)$ network. The first structure is $\text{CNN}+\text{GRU}+{\text{Att}}_{\text{time}}$ which is shown in Fig. 1 and consists of both CNN and GRU networks with an attention mechanism on the time dimension. In the $\text{CNN}+{\text{Att}}_{\text{time}}$ structure, the GRU was removed from the beat encoder, and the attention model was applied to the time dimension. The $\text{CNN}+{\text{Att}}_{\text{freq}}$ structure is similar to $\text{CNN}+{\text{Att}}_{\text{time}}$, except that the attention model was applied to the frequency dimension.

Fig. 7:

Median ± interquartile range estimates of gestational age on NBW (left, green) and LBW (orange, right) individuals. Note that the GA estimates of the LBW are always lower than those of the NBW.

Fig. 8:

Visualization of window level $\left(G_w\left(\cdot ,{\theta }_w\right)\right)$ attention weights. The model assigns lower weights to the low quality segments.

Fig. 9:

Visualization of attention weights using different $G_b\left(\cdot ,{\theta }_w\right)$ structures. Attention weights are shown in red, indicating the importance of different parts of the scalogram in the task of GA estimation. We tested three structures: a) $\text{CNN}+\text{GRU}$ and time attention structure. b) Time attention using just CNN network and c) CNN network and frequency attention.