Automated Assessment of Cardiovascular Sufficiency Using Non-Invasive Physiological Data

Abstract

For fluid resuscitation of critically ill individuals to be effective, it must be well calibrated in terms of timing and dosages of treatments. In current practice, the cardiovascular sufficiency of patients during fluid resuscitation is determined using primarily invasively measured vital signs, including Arterial Pressure and Mixed Venous Oxygen Saturation (SvO2), which may not be available in outside-of-hospital settings, particularly in the field when treating subjects injured in traffic accidents or wounded in combat where only non-invasive monitoring is available to drive care. In this paper, we propose (1) a Machine Learning (ML) approach to estimate the sufficiency utilizing features extracted from non-invasive vital signs and (2) a novel framework to address the detrimental impact of inter-patient diversity on the ability of ML models to generalize well to unseen subjects. Through comprehensive evaluation on the physiological data collected in laboratory animal experiments, we demonstrate that the proposed approaches can achieve competitive performance on new patients using only non-invasive measurements. These characteristics enable effective monitoring of fluid resuscitation in real-world acute settings with limited monitoring resources and can help facilitate broader adoption of ML in this important subfield of healthcare.

Keywords: cardiovascular sufficiency; fluid resuscitation; machine learning; non-invasive monitoring; physiological data.

Figure 1

When the mean of Arterial Pressure and the mean of SvO2 (both invasively measured) are above the target values (dashed lines), the subject is labeled as “sufficient” at the given assessment time, or as “insufficient” otherwise.

Figure 2

The Optimized Aggregation of Predictions framework. From a list of candidate thresholds, the one which maximizes the correlation between the predictions made by the trained model on the validation data normalized with reference to its own personal baseline (Prediction*, denoted by blue) and the aggregated binary predictions made by the same model on the validation data standardized using the normalization factors of different training subjects (Majority Vote %, denoted by red) is chosen to be used for converting the prediction scores to binary predictions on the test data.

Figure 3

The predictions for the test data using the threshold chosen via optimization performed using the validation data as shown in Figure 2.

Figure 4

The mean and the standard error bands of the ROC curves of three different approaches for resuscitation sufficiency prediction. The False Positive Rate (FPR) and the False Negative Rate (FNR) are scaled logarithmically in the middle and right plots to emphasize the performance at the clinically relevant low prediction errors settings.

Figure 5

ROC curves of two example test subjects. The $95\%$ confidence intervals are computed using the Wilson interval scores.

Figure 6

The iso-performance lines and the selected cost-optimal decision thresholds (shown in boxes) corresponding to the three different settings of $\frac{C_{FP}}{C_{FN}}$.

Figure 7

The contingency tables and the p-values for the McNemar’s tests.