Hourly Early Recognition of Sepsis in the Intensive Care Unit Using Variable-length Long Stort Term Memory Recurrent Neural Networks

Rutger van de Leur and Joost Plate
UMC Utrecht


Background Sepsis is a common life-threatening condition, for which early detection and (antibiotic) treatment may prevent mortality. As part of the Physionet 2019 challenge, this study aims to timely predict sepsis at the Intensive Care Unit (ICU).

Methods This retrospective study uses ICU data from two separate hospital systems. As the outcome, sepsis was defined as the combination of clinical suspicion of sepsis (obtaining cultures or the start of antibiotics) and a two-point deterioration in the SOFA score within a 24-hour period. Covariates used are the baseline demographics, vital signs (measured every 2h) and laboratory values. Missing values of the vital signs were inter-and extrapolated using natural cubic splines, whereas for laboratory values the last measured variable was used. To adjust for the differences in sequence length per prediction timestamp, zero padding was utilized. Subsequently, a variable-length long-short term recurrent neural network (LSTM RNN) was fitted for all prediction time stamps. Hyperparamater tuning was performed with Bayesian optimization.

Preliminary results The dataset contained 5000 admissions with a median admission duration of 39 [IQR 23 - 47] hours and a between 0 and 38 repeated measurements per variable. In 179 (5.58 %) admissions sepsis occured. The validation dataset accuracy for this algorithm was 83%.

Conclusion A variable length LSTM-RNN may support the clinician in the timely recognition of sepsis, thereby potentially reducing the mortality at the ICU.