LSTM Modeling of Perinatal Fetal Heart Rate

Abstract

Objective: We use Long Short Term Memory (LSTM) units to model perinatal fetal heart rate. We assess the effect of modeling variants such as dropout, multiple cells/layer, multiple layers on model precision and parameter count. Motivation: Linear models can model short-term signal characteristics. An LSTM should do this as well but in addition, retain important longer term characteristics and do so with a low parameter count model. Methods: Recordings of 115 normal, 115 metabolic acidotic and 44 severely pathological (P) fetuses were detrended and downsampled before LSTM modelling. Results: A single-unit LSTM model produced a succinct (4 state parameters), and expressive model (mean variance accounted for of 87.7%) without overfitting. Larger numbers of units N gave marginal improvement at the cost of much higher parameter count.