This paper presents the first empirical study on using Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs) to classify diagnoses from multivariate time series data in the pediatric intensive care unit (PICU). The authors address the challenges of varying episode lengths, irregular sampling, and missing data in clinical time series. They train an LSTM model to classify 128 diagnoses based on 13 frequently sampled clinical measurements. The study compares the LSTM model to several strong baselines, including a multilayer perceptron (MLP) trained on hand-engineered features. Key contributions include the introduction of target replication, a strategy that replicates targets at each sequence step to improve training and generalization, and the use of auxiliary outputs to reduce overfitting. The results show that the LSTM model outperforms the MLP baseline and its ensembles, demonstrating the effectiveness of LSTMs in learning from complex clinical time series data. The authors also discuss the limitations and future directions, including the need for better data preprocessing and the exploration of other regularization techniques.This paper presents the first empirical study on using Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs) to classify diagnoses from multivariate time series data in the pediatric intensive care unit (PICU). The authors address the challenges of varying episode lengths, irregular sampling, and missing data in clinical time series. They train an LSTM model to classify 128 diagnoses based on 13 frequently sampled clinical measurements. The study compares the LSTM model to several strong baselines, including a multilayer perceptron (MLP) trained on hand-engineered features. Key contributions include the introduction of target replication, a strategy that replicates targets at each sequence step to improve training and generalization, and the use of auxiliary outputs to reduce overfitting. The results show that the LSTM model outperforms the MLP baseline and its ensembles, demonstrating the effectiveness of LSTMs in learning from complex clinical time series data. The authors also discuss the limitations and future directions, including the need for better data preprocessing and the exploration of other regularization techniques.