Objectives Early detection of sepsis reduces mortality as antibiotics can be administered earlier. A recurrent neural network was developed to identify the early onset of sepsis in ICU patients for the PhysioNet/CinC Challenge 2019. Deep learning models require the selection of many hyperparameters, such as the learning rate and number of layers, which often is done somewhat arbitrarily. To address this, hyperparameters for the network were chosen with the Optometrist algorithm, which combines stochastic perturbations with human choice.

Methods ICU stays were released for the challenge, each with 40 variables including vital sign measurements and demographic information. 90% of the initial 5000 patients were used to train the model, and the other 10% for the development set. The final model was validated against an external test set. Patient data was normalized and fed into a recurrent neural network, followed by dense layers.

Hyperparameter optimization began by training two sets of hyperparameters. The best performing model was selected by comparing performance metrics in a side-by-side comparison. Another hyperparameter set was chosen by randomly increasing/decreasing the best hyperparameters. After training, these results were compared with the best model. This process was then repeated, with the new hyperparameters replacing the best hyperparameters if they showed better performance. If the new hyperparameters were chosen the variance of the next set of hyperparameters increased, otherwise the variance decreased.

Results There were 14 different hyperparameters that were optimized. With the optometrist algorithm, an exhaustive search was not required to find the optimum hyperparameters, as each comparison took seconds to complete. The final set of hyperparameters had 1 LSTM layer with 79 neurons, 5 dense dense layers with 225 neurons in each layer. The competition utility score achieved on the test set was -0.011.