Using tensor-reduced physiological signals to predict sepsis trajectory
Date:
Background: The quick sequential organ failure assessment (qSOFA) scoring system, defined by Sepsis-3 in 2016, is a method to identify patients at risk to progress to poor outcomes related to sepsis using variables that can be collected at the bedside. Methods: Our group used the learning methods Support Vector Machine, Learning Using Concave and Convex Kernels, and Random Forest to predict an increase in qSOFA using electronic health record (EHR) data. We used Random Forest for further analysis using waveform features extracted from electrocardiogram and arterial line. Results: Our group found that although a Random Forest model trained on only EHR data can create adequate predictions in a 6-hour time frame (AUROC 0.781 ± 0.113), one trained on waveform features can create a similar prediction (AUROC 0.739 ± 0.118). A model trained on waveform features is further improved when the data are structured as a tensor, and tensor decomposition via Canonical Polyadic / Parallel Factors with Alternating Least Squares (CP-ALS) is used to reduce the feature space (AUROC 0.753 ± 0.116). Conclusion: A waveform-informed model offers distinct advantages over an EHR data-informed model, despite experiencing a reduction in performance. The first is that predictions can be made continuously in real-time, and second is that these predictions would not be limited by the availability of EHR data. Additionally, structuring the waveform features as a tensor conserves structural and temporal information that would otherwise be lost if the data were presented as flat vectors.
