14 - Towards automated detection and quantification of paroxysmal sympathetic hyperactivity after traumatic brain injury: contextualized physiologic data visualization

Abstract Text:

Introduction: Paroxysmal sympathetic hyperactivity (PSH) is a clinically important syndrome of episodic extreme physiologic derangements that occurs in a subset of patients following traumatic brain injury (TBI). Its occurrence has been associated with worse outcomes even when controlling for initial injury severity, suggesting that it may represent a targetable mechanism of secondary injury. However, diagnosis is difficult and often made late, after the exclusion of other potential causes, limiting clinicians’ abilities to mitigate its detrimental effects. The consensus-based PSH-Assessment Measure (PSH-AM) has led to the standardization of terminology, diagnostic criteria, and rudimentary quantification of PSH but has a number of limitations. Through collaboration between neurointensivists and data scientists, we aim to develop an automated PSH quantification tool. Here, we evaluate proof of concept with data visualization and model training from prototypical patient data.

Methods: Adult critically ill acute TBI patients admitted for at least 14 days with high resolution physiologic data recordings available from the R Adams Cowley Shock Trauma Center between January 2016 and July 2018 were identified via the institutional trauma registry. Additional data collected included medication administration records, demographic, clinical, and radiographic data extracted from the electronic medical record (EMR), and daily PSH-AM scores (retrospectively allocated by review of the EMR). Raw and transformed time series data were integrated into an existing open-source data visualization and annotation tool, Auton Universal Viewer (AUViewer, https://auviewer.readthedocs.io/en/latest/). Data was checked for quality, artifacts removed, and optimal viewing frames explored to enable PSH event annotation in several prototypical patients with versus without PSH.

Results: We explore data visualizations for several of 221 eligible patients. For each, we present flexible and customizable time series visualizations of heart rate, heart rate variability, respiratory rate, blood pressure, temperature, daily PSH-AM scores, novel instantaneous modified PSH-AM scores, and relevant medication dosage and infusion rates. We demonstrate the generation and refinement of PSH labeling rules using a combination of anomaly detection and human-in-the-loop machine learning from this data. We also demonstrate a positive qualitatively comparison of PSH detection using this versus an EMR review strategy.

Conclusions: Physiologic data visualization-in-context may enable a human-in-the-loop machine learning and anomaly detection strategy for the iterative development of an automatic quantitative “PSH index.” Initial PSH labeling functions derived from prototypical examples shown here will undergo refinement as they are applied to the full retrospective cohort and validation in a prospective manner. Future work will also assess the clinical utility of the PSH index as a diagnostic, monitoring, and predictive biomarker in future studies targeting the sympathetic nervous system in the management of TBI.

Keywords: Paroxysmal sympathetic hyperactivity, dysautonomia, physiology, critical care, machine learning, data visualization, big data, biomarker