University of Massachusetts Amherst

Early sepsis detection is an important topic in the medical domain due to its high incidence expectancy and low mortality rate. The recent advancements in deep learning frameworks help in modeling such multi-resolution time-series problems more efficiently. For Physionet/Computing-in-Cardiology 2019 challenge, we propose a multi-task deep learning framework that utilizes dot attention. We formulate the problem of early sepsis detection in three parts: first, efficient imputation of missing values for different signals, second, predicting the occurrence of sepsis and third, predicting sepsis onset timestamp for the patient. Currently, we focussed on the second and third task in order to develop and test our modelâ€™s architecture. For missing data problem, we did mean imputation wherever the values were missing. The mean-imputed time series was fed to a recurrent neural network (RNN): Bidirectional long short term memory (BiLSTM). The dot attention mechanism is applied to the output of BiLSTM network which identifies the important sub-sequences of such time series data and weighs them accordingly to get the final vector representation. This representation is used for predicting whether the patient will have sepsis or not and is simultaneously used to predict the timestamp of its onset. The weighted sum of losses for these two tasks become the total loss for our model and is used to train the parameters of the model. Currently, linear interpolation is used for generating the probability of sepsis at each time stamp. Our further plans include using smarter imputation techniques such as temporal belief memory and creating more sophisticated sequence network. For evaluation, we separated data into training and testing in the ratio of 80:20. Our current approach achieved a utility score of 0.485 with minimal hyperparameter tuning. The utility score would improve further after using improved imputation, interpolation strategies and hyperparameter tuning of the network.