Generalizable deep temporal models for predicting episodes of sudden hypotension in critically ill patients: a personalized approach

Figures

• 

  Figure 1. Experimental workflow and study design. Data from both the KGH and MIMIC-III databases are annotated and processed before being partitioned into three sets that are used for training, validation, and testing. Initial model development is performed using data from KGH exclusively. The models are then adapted via transfer learning using MIMIC-III data. All models are evaluated using held-out internal and external test sets.

• 

  Figure 2. Comparing a patient specific definition (left) of AHE to the threshold based definition (right). Our patient specific definition is based on the use of two moving averages of MAP recordings in which the outcome of interest is a 20% drop in the averages. The conventional definition of AHE is based on a hard threshold such as 60 mmHg as highlighted in red. In this case, the threshold based definition would not annotate an AHE in this patient recording as a MAP below 60 mmHg is not sustained.

• 

  Figure 3. Comparison of AUROC between the logistic regression models, SVM models, and deep models on both the internal and external test sets at each investigated observation and gap length.

• 

  Figure 4. Comparison of ROC curves and corresponding 95% confidence intervals of models trained to make predictions using a 60 min observation window at a gap length of 10 (left), 20 (center), and 30 (right) minutes. Red curves denote performance of the LSTM model on the internal test set, green the LSTM model on the external test set, blue the GRU-D model on the internal test set, and purple the GRU-D model on the external test set.

• 

  Figure 5. An example of an identified event in a patient recording. The annotated 20% drop is representative of the difference between a 60 min moving average (orange) and a 5 min moving average (green) of MAP. The relationship between the intervals of time that represent the observation window (highlighted green) and gap length (highlighted red) are also depicted relative to the identified drop in MAP.

• 

  Figure 6. The architecture of the LSTM network. A window of physiological data containing MAP, HR, and SPO2% recordings is processed through 2 LSTM layers with the output of the last LSTM layer being merged in a final layer to make a binary prediction. LSTM units in the first layer are shown with different colors (unit 1: lighter red through unit 60: darker red).