An investigation of patterns in hemodynamic data indicative of impending hypotension in intensive care

Abstract

Background: In the intensive care unit (ICU), clinical staff must stay vigilant to promptly detect and treat hypotensive episodes (HEs). Given the stressful context of busy ICUs, an automated hypotensive risk stratifier can help ICU clinicians focus care and resources by prospectively identifying patients at increased risk of impending HEs. The objective of this study was to investigate the possible existence of discriminatory patterns in hemodynamic data that can be indicative of future hypotensive risk.

Methods: Given the complexity and heterogeneity of ICU data, a machine learning approach was used in this study. Time series of minute-by-minute measures of mean arterial blood pressure, heart rate, pulse pressure, and relative cardiac output from 1,311 records from the MIMIC II Database were used. An HE was defined as a 30-minute period during which the mean arterial pressure was below 60 mmHg for at least 90% of the time. Features extracted from the hemodynamic data during an observation period of either 30 or 60 minutes were analyzed to predict the occurrence of HEs 1 or 2 hours into the future. Artificial neural networks (ANNs) were trained for binary classification (normotensive vs. hypotensive) and regression (estimation of future mean blood pressure).

Results: The ANNs were successfully trained to discriminate patterns in the multidimensional hemodynamic data that were predictive of future HEs. The best overall binary classification performance resulted in a mean area under ROC curve of 0.918, a sensitivity of 0.826, and a specificity of 0.859. Predicting further into the future resulted in poorer performance, whereas observation duration minimally affected performance. The low prevalence of HEs led to poor positive predictive values. In regression, the best mean absolute error was 9.67%.

Conclusions: The promising pattern recognition performance demonstrates the existence of discriminatory patterns in hemodynamic data that can indicate impending hypotension. The poor PPVs discourage a direct HE predictor, but a hypotensive risk stratifier based on the pattern recognition algorithms of this study would be of significant clinical value in busy ICU environments.

Figure 1

A graphical illustration of the observation window, gap, and target window with respect to prediction time.

Figure 2

An example of minute-by-minute continuous prediction with HEs. The gap and observation window sizes are 1 and 0.5 hour, respectively. The top plot shows the true and predicted (from regression) MAPs, and the dashed line at 60 mmHg represents the HE threshold. The bottom plot shows the corresponding posterior probabilities of hypotension from the binary classifier. The results from both single and multiple compilation are shown. The periods of no prediction are due to inadequate input data. To facilitate comparison in the top plot, the predicted values were shifted to remove the gap. This implies that the predictions were actually generated 1 hour in advance. However, the bottom plot was not temporally aligned to preserve the predictive nature.

Figure 3

An example of minute-by-minute continuous prediction with no HE. For details, please see the Figure 2 caption.

Figure 4

A visual illustration of PCA loading on the first 19 principal components that captured 90% of the total variance in the entire multiple compilation data. Only the statistical and clinical features are shown; see Figure 5 for the cross-correlation and wavelet features. The principal components are sorted in decreasing order of eigenvalue, with PC1 associated with the largest eigenvalue. The gap and observation window were 1 hour and 30 minutes long, respectively. The size of each rectangle is directly proportional to the magnitude of the loading. Filled and unfilled rectangles correspond to positive and negative loading coefficients, respectively.

Figure 5

Same visualization of PCA loading as Figure 4 for the cross-correlation and wavelet features. See Figure 4 caption for more details.