Fig 1.
LSTMs incorporating attention can model multiple clinical targets.
(A) The schematic of a basic attention-LSTM sequence to sequence architecture. The input variable-level attention layer (highlighted in red) is passed to a recurrent neural network LSTM cell which then sends both an encoded cell state vector and an encoded output to the next time point, the latter of which is used to make a prediction for the target at each time point. (B-D) Three AUROC curves demonstrate this approach’s ability to model same-day myocardial ischemia (AUC 0.834), sepsis (AUC 0.952), and vancomycin administration (AUC 0.904) in the ICU. We note that since predictors and targets are drawn from the same time window in this formulation of the model, this is a modeling rather than predictive task.
Fig 2.
Next-day predictive performance varied depending on the target and location in the time series.
(A) Sepsis (next-day AUC 0.876) shows a clear trend towards improved predictive performance later in the course, while (B) vancomycin administration (next-day AUC 0.833) is predicted similarly well over most time periods. (C) Myocardial infarction (next-day AUC 0.823) is rare in later time periods.
Fig 3.
Attention heat maps averaged over all patients with daily attention and features with highest mean activation per day.
As described in “Extracting clinical information from average attention” in Methods, heatmaps are second order tensors produced by averaging over all patients. These patient-averaged heatmaps contain activations for each feature at each time step. One-dimensional feature-averaged activations are plotted above each heatmap and reveal days which were most highly attended to by the model. Lastly, each panel contains the activations for the feature which was most heavily attended to at every time step. The patient-averaged attention across all patients with (A) myocardial ischemia, (B) sepsis, and (C) vancomycin, demonstrates emphasis on early days, reflecting the fact that modeling initiation days are critical for making accurate predictions over the time series.
Fig 4.
Personalized attention maps for three patients whose ICU courses were positive for myocardial ischemia, sepsis, and vancomycin.
Normalized input features in (A) are fed to the attention mechanism whose activations are visualized in (B). Ground truth labels and predictions for all target endpoints are labeled in (C). The days preceding the day of the ICU event, features with the highest time-relative activation on that day were identified in (D). The attention mechanism on the day of MI revealed a focus on nitroglycerin administration. For sepsis, administration of ranitidine and ceftriaxone were highly attended to in the days preceding the ICU event. Finally, in the case of vancomycin, the model attended heavily to labetalol, phenylepherine, and cefepime leading up to vancomycin requirement.