Incorporating Pathophysiological Knowledge into a Time-Aware Long-Short Term Memory for the Early Prediction of Sepsis

Marcus Vollmer1, Christian F. Luz2, Philipp Sodmann1, Maarten W. N. Nijsten3, Bhanu Sinha2, Sven-Olaf Kuhn4
1Institute of Bioinformatics, University Medicine Greifswald, 2Department of Medical Microbiology, University Medical Center Groningen, 3Department of Critical Care, University Medical Center Groningen, 4Department of Anesthesiology and Intensive Care Medicine, University Medicine Greifswald


Motivation: This contribution relates to the PhysioNet/CinC Challenge 2019 on real-time early detection of sepsis from bedside monitoring and laboratory parameters. Accounting for complex clinical dynamics in sepsis patients while aiming at an automated analysis of hourly (non-)validated data is challenging. The algorithm has to deal with imprecise, incorrect and incomplete data in addition to being time aware.

Methods: We aim to use a Time-Aware Long Short-Term Memory (T-LSTM), a recurrent neural network for handling irregular time intervals, in longitudinal patient records. Our Python-implemented open-source algorithm aims at the integration of laboratory parameters, continuously measured vital signs, and pathophysiological knowledge in the architecture of the T-LSTM model. Missing data in hourly-measured variables is handled with multiple imputations using credible intervals. The T-LSTM is trained on an 80\% validation split of 40,336 ICU patients. The sepsis prediction will be externally validated on the hidden test set of the 2019 PhysioNet/CinC Challenge. The performance is based on the utility functions that reward a prediction between 12 hours before and 3 hours after the sepsis onset (as defined by the Sepsis-3 guidelines) with a defined optimum at 6 hours before. Furthermore, results will be compared with an ordinary LSTM.

Results: A T-LSTM was developed for the early prediction of sepsis in ICU patients processing hourly measurements of vital signs and irregular time intervals of laboratory parameters together with medical knowledge. A normalized utility score of 0.30 can be reported for a LSTM submitted during the unofficial phase evaluated with the training data set. Further preliminary results are reported in the table below (see PDF abstract).