This study aims to develop, as a proof of concept, a previously described (10–13) deep learning methodology, namely a long short-term memory (LSTM) recurrent neural network (RNN) to continuously assess an individual child’s ROM throughout their ICU stay as a proxy measure of SOI. Deep learning methods combine variables in many more different ways than logistic regression, giving rise to many more coefficients or weights than the number of input variables, enabling them to capture more complex interactions among inputs than those captured by simpler algorithms such as logistic regression. Improvements in regularization techniques, including L1 and L2 constraints, dropout techniques, initial learning rate, and learning rate decay, enable deep learning models to manage hundreds of inputs while improving their accuracy.

RNNs are specifically designed to process sequential data. The “recurrent” architecture allows the integration of information from previous timesteps with newly acquired data to update its risk assessment, making the model dynamic instead of static. RNNs analyze all available data with neither preconception about which measures may be important in determining a patient’s clinical status nor the need to engineer features specific to a given clinical condition.

Reference link- https://journals.lww.com/pccmjournal/fulltext/2021/06000/continuous_prediction_of_mortality_in_the_picu__a.2.aspx

Author