LSTM as ElectroHysteroGram Signal Forecasting Method

In this work, the Long Short-Term Memory architecture has been exploited for ElectroHysteroGram signal forecasting over a given horizon. It is an important step allowing the signal’s future values anticipation, for labour prediction in pregnant women. Pytorch framework was used to develop the proposing forecasting model. Term-Preterm ElectroHysteroGram database signals were used to evaluate the performance of the model. After several trainings on different datasets, it was found that the best model is composed of 50 neurons per hidden layer and five hidden dense layers. It also allows the signal’s forecasting from any position on it. Moreover, the total duration of the considered dataset can be considered as forecasting horizon. This makes it more efficient than the N-BEATS model developed in our previous work for the same purpose. But it is necessary to consider a training batch_size of 128 and the appropriate backcast window as hyperparameters, to obtain the best training and validation loss curves. So, the signal’s trend and seasonality are followed by the model, after the forecasted and current values comparison. Nevertheless, an overall average is observed, which does not allow the extreme values’ captation by the model. This work represents an important step for a real time labour prediction in pregnant women. It should come before the classification stage on which most of the literature models have focused. In this way, decision making and access to hospital delays can be reduced in low-and middle-income countries.