Characterization of multi-view hemodynamic data by learning mixtures of multi-output regressors

Abstract

Introduction: We investigate whether central venous pressure (CVP) pressure waveform signal can be informative in detection of slow bleeding in post-surgical patients. We apply a novel machine learning method to analyze CVP datasets to characterize bleeding in a porcine model of fixed rate blood loss.

Methods: Thirty-eight pigs were anesthetized, instrumented with catheters, kept stable for 30 minutes, and bled at a constant rate of 20ml/min to mean arterial pressure of 30 mmHg. CVP waveforms were extracted from inspiration and expiration phases of respiration and statistically featurized. The proposed machine learning method, Canonical Least Squares (CLS) clustering, identifies correlation structures that differ between subsets of observations. We extend it to supervised classification. Both clustering and classification methods yield human-interpretable models that reflect distinctive patterns of correlations within CVP waveforms.

Results: We conducted three experiments to discover structure in the physiological response to bleeding. First, we clustered respiration cycles with full knowledge of blood loss. The color-coded cluster assignments are shown in the Figure 1. They are consistent with escalation of bleeding. Second, we deployed clustering on only CVP features without blood loss. Temporal structure was complemented with some subject-specific clusters (Fig 2). Third, we ran CLS classification to decide whether an observation came from before or after the onset of bleeding (performance shown in the Table 1).

Conclusions: Our results show that the CVP waveforms carry information about the physiologic status of the subject and indicate the amount of blood lost. Clusters still correspond to bleeding status even when no information about bleeding was available to the algorithm. CLS clustering enables a detailed yet interpretable view of discovered structures in complex waveform data.