Data-driven patient clustering represents a promising emerging approach to break down heterogeneity in TBI based on underlying mechanistic processes rather than relying on clinical constructs[26–28]. Using unsupervised K-means clustering and trajectory group analysis, we discovered four distinct CFS trend patterns during the first two weeks of acute TBI hospitalization. Beyond its relevance toward understanding the evolution of the specific clinical syndrome of PSH, this work represents a first step toward the identification of naturally-occuring autonomic endotypes in critically ill TBI patients. As such, the observed trajectory patterns demonstrate more nuance and divergence than those comparing clinically diagnosed PSH cases and controls (Fig. 3 vs Supplementary Material Fig. 1) and are worthy of further exploration.
Observed trajectory patterns differ in both initial CFS scores and evolution thereafter. We note that the trends following the initial days are more critical in predicting PSH diagnosis and DLT scores at day 14 than the initial CFS scores. Specifically, if the CFS trend does not decrease after 4 or 5 days, there is a higher likelihood of a positive PSH diagnosis. In contrast, a downward trend in CFS after the initial period correlates with a lower probability of PSH. The analysis suggests that CFS trends hold more promise as early predictors of PSH compared to initial CFS scores. Future work is warranted to identify the precise timing and trend features for the earliest reliable prediction of PSH. Correlation analysis between trajectory groups and outcome variables beyond PSH diagnosis demonstrated significant relationships with ICU LOS and number of days on mechanical ventilation. Therefore, early prediction and successful management of PSH could alter these outcomes, improving healthcare resource utilization and reducing clinical complications associated with longer ICU LOS and duration of mechanical ventilation. We did not observe significant relationships between trajectory group membership and other outcomes including mortality and hospital discharge GCS. We believe that this lack of observed effect may have been due to our relatively small, homogenous sample size of severe TBI patients who were hospitalized for at least 14 days, with selection bias leading to low mortality, and lack of long-term follow-up outcome data leading to fairly homogenous and crude outcomes. It remains unknown whether longer term functional outcomes differ according to sympathetic activation trajectory groups.
Exploring the relationships between admission characteristics and physiologic trajectory groups sheds light on potential predictors of PSH and naturally-occurring phenotypes of sympathetic nervous system activation following acute TBI. Univariate screening identified a subset of baseline and early admission variables significantly associated with trajectory group membership, including age, BMI, tGCS, mGCS, and use of invasive ICP monitoring. In the subsequent fitting of the multivariate multinomial logistic regression model, only two variables were found to be relevant: age and mGCS. Specifically, the odds of belonging to high CFS groups (3 & 4) are decreased by 3% and 4% for each increase of one year in age after controlling for mGCS score. This observation aligns with previous literature suggesting that younger age is associated with a higher risk of PSH in adults[17, 29, 30]. Interestingly, the opposite effect has been observed in pediatric populations, where older age is associated with greater risk for PSH[31]. Rather than a simple ordinal risk factor, age may therefore be an important grouping factor for physiologic trajectories across the entire span of neurodevelopment and natural aging. We also found that for each unit decrease in mGCS, the odds of belonging to group 4 (persistently high CFS and increased risk for PSH) increased by 9%. This is consistent with published literature suggesting that lower initial GCS increases the risk for PSH[32].
Given that the CFS score is composed of clinical and physiologic markers of sympathetic nervous system activation, CFS trajectory groups may provide insight beyond prediction of the syndrome of PSH, extending into the drivers and evolution of post-TBI dysautonomia, more generally. Following TBI, sympathetic activation has been associated with higher injury severity, increased mortality, and mechanisms of secondary injury including inflammation, coagulopathy, endothelial dysfunction, and glymphatic system dysfunction[33–36]. Further, there is evidence to suggest that interventions that reduce sympathetic activation and its end-organ consequences block these effects[36–38]. However, many features related to sympathetic activation including its timing, magnitude, persistence and pharmacologic modulation, may contribute to its effects on secondary injury mechanisms and outcomes. As a persistent but paroxysmal phenomenon, it is unclear whether PSH specifically contributes to these sympathetically mediated effects, or whether alternative autonomic phenotypes should be considered as targets for interventions to improve outcomes. Nonetheless, the PSH-AM score provides a framework for quantifying sympathetic hyperactivity via the CFS score and may serve as a starting point for phenotyping post-TBI dysautonomia in a standard way.
This study represents preliminary work with a number of limitations beyond those already discussed. First, CFS trends were derived from retrospectively allocated CFS scores based on EHR review of vital signs and clinical documentation. Prospective documentation of PSH episodes and analysis using raw vital signs may provide more accurate and nuanced physiologic information. We used a convenience sample of critically ill TBI patients who were scored according to the PSH-AM as part of prior work[7, 39, 40]; as a whole our included cohort was younger with more severe injuries compared to the held out eligible patients (Supplementary Material Table 1) and therefore may have been at higher risk for post-TBI dysautonomia.
Larger, multi-center prospective studies are needed to confirm and expand upon these findings. Larger datasets will generate more power to identify baseline, clinical, and outcome variables associated with each trajectory group. For example, we demonstrated trend-level associations between initial CT radiographic features and trajectory group membership, which could become significant with more statistical power. Stronger phenotypic descriptions of the trajectory groups may further suggest heterogeneous mechanistic targets for autonomic nervous system targeted therapies in traumatic brain injury.