Time Series Prediction of Humidity in Neonatal Incubators Using NARX and LSTM Neural Network Architectures

Abstract

Humidity control is a key factor in neonatal care, particularly inside incubators, where it is essential to maintain stable environmental conditions. This study focuses on two humidity forecasting approaches: the NARX model, that is, a nonlinear autoregressive network with exogenous variables, and the LSTM architecture, a type of neural network known as Long Short Term Memory, both of which are well suited for time series modeling. The main objective is to evaluate and compare the predictive performance of these two models using four statistical indicators: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the coefficient of determination <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathrm{R}^{2})$</tex>. The data used were collected from a neonatal incubator at the LAPER laboratory, consisting of continuous humidity recordings over a six hour period, resulting in a total of 2096 samples. After training and validating the models, the following results were obtained: LSTM (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{MSE}=0.010, \text{RMSE}=0.103, \text{MAE}=0.067, \mathrm{R}^{2}=0.989$</tex>) and NARX (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{MSE}=0.317, \text{RMSE}=0.563, \text{MAE}=0.392, \mathrm{R}^{2}=0.990$</tex>). Although the NARX model shows a slightly higher coefficient of determination, the LSTM model yields significantly lower error values, indicating a better ability to capture subtle variations in humidity. These results highlight the robustness of the LSTM model and confirm its effectiveness as a reliable and accurate tool for humidity regulation in sensitive medical environments such as neonatal incubators.