A machine-learning based analysis for the recognition of progressive central hypovolemia

Traditional patient monitoring during surgery includes heart rate (HR), blood pressure (BP) and peripheral oxygen saturation. However, their use as predictors for central hypovolemia is limited, which may lead to cerebral hypoperfusion. The aim of this study was to develop a monitoring model that can indicate a decrease in central blood volume (CBV) at an early stage.

Twenty-eight healthy subjects (aged 18-50 years) were included. Lower body negative pressure (-50 mmHg) was applied to induce central hypovolemia until the onset of pre-syncope. Ten beat-to-beat and four discrete parameters were measured, normalized, and filtered with a 30 s moving window. Time to pre-syncope was scaled from 100%-0%. A total of 100 neural networks with 5, 10, 15, 20, or 25 neurons in their respective hidden layer were trained by 10, 20, 40, 80, 160, or 320 iterations to predict time to pre-syncope for each subject. The network with the lowest average slope of a fitted line over all subjects was chosen as optimal.

The optimal generalized model consisted of 10 hidden neurons, trained using 80 iterations. The slope of the fitted line on the average prediction was  -0.64 (SD 0.35). The model recognizes in 75% of the subjects the need for intervention at  >200 s before pre-syncope.

We developed a neural network based on a set of physiological variables, which indicates a decrease in CBV even in the absence of HR and BP changes. This should allow timely intervention and prevent the development of symptomatic cerebral hypoperfusion.