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

Tomas Vicar1, Jakub Hejc1, Petra Novotna1, Marina Ronzhina1, Radovan Smisek2
1Department of Biomedical Engineering, Brno University of Technology, 2Brno University of Technology


In this entry, the first version of the methodology for early prediction of sepsis from clinical data is presented. Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection and early prediction of this state at intensive care units (ICUs) could be crucial for patient treatment. For sepsis prediction from provided time series data, a deep-learning method was chosen. Long Short-Term Memory networks (LSTM) are the most suitable methods for time series data. However, a large number of missing values is causing difficulties in the direct use of LSTM on provided clinical data. LSTM network is not applicable to sparse data. Also, simple interpolation introduces distortion of the data. Therefore, dealing with missing values is the biggest challenge of the methodology. The proposed method consists of several steps. 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 classes automatically weighted by the occurrence of the feature in each batch. This leads to more stable network training. The proposed network consists of 2 bidirectional LSTM layers followed by 8 fully connected layers. The output of the network is sepsis prediction value at each point in time. The version of the implemented method was submitted to the system on Friday, April the 5th. Unfortunately, at the moment of abstract submission, official utility score values were not published yet. In internal testing, the normalized utility score of 0.7351 was reached, evaluated by 10-fold cross-validation on the provided training dataset.