Sepsis Detection in Sparse Clinical Data Using Long Short-Term Memory Network with Dice Loss

Abstract

This paper aims to present a methodology for sepsis prediction from clinical time-series data. Sepsis is one of the most threatening states which could occur while treating a patient at the intensive care unit. Therefore its prediction could significantly improve the quality of the patient treatment.In this work, we address the problem of sepsis prediction with Long Short-Term Memory (LSTM) network with specialized deep architecture with residual connections. The output of the network is sepsis prediction score at each point in time.Feature normalization into the fixed range of values is applied including replacing missing values with numerical representation from outside the normalized range. Therefore, the LSTM network is able to include missing values in the learning process. Also, the rarity of sepsis occurrence in the provided dataset is a challenging problem. This problem is addressed by the application of dice loss providing automatically weighted classes by the occurrence of the feature.The proposed method leads to 0.281 normalized utility score on the full test set as the best official Phys-ioNet/Computing in Cardiology (CinC) Challenge 2019 entry of ECGuru10 team.

Comparative Study of Machine Learning Models for Onset Sepsis Prediction

“…Vicar et al [9] proposed a model based on Long Short-Term Memory (LSTM) network consists of 7 blocks, where each block consists of LSTM layer followed by 3 fully connected layers. Dataset unbalancing was handled by a specialized lost function, which can automatically set the weights appropriately for each of the classes [10], while missing values were handled by replacement of missing values with the average value in most of the cases.…”

ECG Abnormalities Recognition Using Convolutional Network With Global Skip Connections and Custom Loss Function

“…Original 2D convolutional filters were replaced by its 1D equivalents. In order to address vanishing gradient problem efficiently, skip connections across residual layers were extended by global skip connection [4] providing a direct shortcut to the model input layer (Input Gate in Figure 1). ResNet based feature extractor consists of 6 residual blocks each of which contains 3 convolutional layers (k = 3).…”

Two-Stage Monitoring of Patients in Intensive Care Unit for Sepsis Prediction Using Non-Overfitted Machine Learning Models

“…Nemati et al successfully used random subsampling to train deep cancer subtype classifier [6]. Vicar et al used special cost function-Generalized Dice Loss [7]. Sweetly et al created 54 datasets using the same sepsis data and different non-sepsis data records [8].…”