Sepsis Prediction Using Advanced Imputation and Long Short Term Memory Networks

Jonathan Rubin, Yale Chang, Gregory Boverman, Shruti Vij, Asif Rahman, Annamalai Natarajan, Saman Parvaneh
Philips Research North America


Background:  Sepsis is a major health crisis and its prediction is of importance for providing timely intervention. The goal of 2019 PhysioNet/CinC Challenge is to develop algorithms to predict sepsis. In this article, an advanced imputation algorithm and long short-term memory network (LSTM) was used to predict sepsis up to 6 hours before its occurrence.    Method:  The training dataset, consisting of 5000 patients, was split randomly into five folds. The number of sepsis cases was not significantly different between folds (p=0.84 using Chi-squared test). Furthermore, there was no statistically significant difference between non-sepsis and sepsis groups for length of data, duration of non-sepsis and sepsis as determined by one-way ANOVA (p>0.127). Three folds (non-sepsis: 2,810, sepsis: 168), one fold (non-sepsis: 954, sepsis: 58), and one fold (non-sepsis: 957, sepsis: 53) were used for in-house training, validation, and test sets, respectively.  

Time-series of vital signs, laboratory values and patient demographics were fed into an LSTM network that was trained to predict outcome sequences. Missing feature values were learned on the training dataset and used to perform imputation. A novel utility function designed by challenge organizers was used for assessing the algorithm’s performance.

Results:  Utility function scores on the in-house validation and test set were 0.59 and 0.52, respectively. Furthermore, AUROC, AUPRC, accuracy and F scores were 0.89, 0.17, 0.95, 0.27 on the validation set and 0.90, 0.12, 0.94, 0.21 on the test set.   

Discussion:   We proposed an advanced imputation method together with LSTMs for early detection of sepsis. Our current results are calculated on an initial cohort of 5000 patients. In future, we will train our algorithm on larger available datasets. Initial results of the method are encouraging. However, further effort is required to improve the algorithm’s performance along with exploring the clinical importance of variables used within the model.