This paper presents the first empirical study using Long Short-Term Memory (LSTM) recurrent neural networks (RNNs) to classify diagnoses based on multivariate time series data from pediatric intensive care unit (PICU) episodes. The study addresses the challenge of analyzing clinical data, which is often irregularly sampled and contains missing values. The authors propose a method to effectively model these time series using LSTMs, which are capable of capturing long-range dependencies and varying sequence lengths. They demonstrate that LSTMs can outperform traditional baselines, including a multilayer perceptron (MLP) trained on hand-engineered features, when trained only on raw time series data. The study uses a dataset of 10,401 PICU episodes, each containing 13 clinical measurements. The goal is to classify each episode with one or more diagnoses from a set of 128 common diagnosis codes. The authors introduce a target replication strategy, inspired by deep supervision techniques, which helps in training the LSTM by replicating targets at each sequence step. This strategy, combined with dropout regularization, improves model performance and reduces overfitting. Additionally, the use of auxiliary outputs from the patient's chart further enhances the model's ability to generalize. The experiments show that the LSTM with target replication and dropout outperforms the MLP on multiple metrics, including micro AUC, macro AUC, and precision at 10. The results indicate that LSTMs can effectively classify diagnoses from clinical time series data, even when the data is irregular and incomplete. The study also highlights the importance of regularization techniques in improving model performance on limited datasets. The findings suggest that LSTMs are a promising approach for diagnosing critical care patients based on clinical time series data, and further research is needed to explore their application in predicting future conditions and treatment responses.This paper presents the first empirical study using Long Short-Term Memory (LSTM) recurrent neural networks (RNNs) to classify diagnoses based on multivariate time series data from pediatric intensive care unit (PICU) episodes. The study addresses the challenge of analyzing clinical data, which is often irregularly sampled and contains missing values. The authors propose a method to effectively model these time series using LSTMs, which are capable of capturing long-range dependencies and varying sequence lengths. They demonstrate that LSTMs can outperform traditional baselines, including a multilayer perceptron (MLP) trained on hand-engineered features, when trained only on raw time series data. The study uses a dataset of 10,401 PICU episodes, each containing 13 clinical measurements. The goal is to classify each episode with one or more diagnoses from a set of 128 common diagnosis codes. The authors introduce a target replication strategy, inspired by deep supervision techniques, which helps in training the LSTM by replicating targets at each sequence step. This strategy, combined with dropout regularization, improves model performance and reduces overfitting. Additionally, the use of auxiliary outputs from the patient's chart further enhances the model's ability to generalize. The experiments show that the LSTM with target replication and dropout outperforms the MLP on multiple metrics, including micro AUC, macro AUC, and precision at 10. The results indicate that LSTMs can effectively classify diagnoses from clinical time series data, even when the data is irregular and incomplete. The study also highlights the importance of regularization techniques in improving model performance on limited datasets. The findings suggest that LSTMs are a promising approach for diagnosing critical care patients based on clinical time series data, and further research is needed to explore their application in predicting future conditions and treatment responses.