In this study, we present a multi-faceted approach to examine dynamic physical, laboratory, and radiographic assessments over time in the first several days after large MCA stroke. We curated a unique dataset of 635 patients with granular, repeatedly measured data. We investigated biomarker courses over time in relation to three life-threatening space occupying mass effect outcomes.
In our population we observed small but noticeable increases in WBC and temperature that preceded radiographic and clinical deterioration due to space-occupying mass effect. These findings were revealed by looking at the variables backward from outcome occurrence and were not obvious in conventional forward-looking plots from the time of admission. While the magnitude was small, WBC and temperature trajectories were consistent in directionality and progressively more pronounced for the three outcomes, suggesting that these variables may be markers of increasing severity of space-occupying mass effect. The etiology of the increase in WBC and temperature leading up to radiographic and clinical indicators of space-occupying mass effect findings is unknown. However, potential reasons may be that increased WBC and temperature are reflective of either an increased inflammatory response due to edema itself or that concurrent infection may in fact worsen edema. We suspect that the observation that sodium mildly increases prior to substantial PGS and DHC reinforces the validity of the trajectory modeling approach, as it is often a treatment target in patients with increasing mass effect. We also observed that heart rate declined over the 24 hours prior to DHC. Physiologically, extreme shifts of brain tissue and herniation, which typically occur after MLS \(\ge\)5mm, have been associated with classic syndromes including Cushing’s Triad, comprised of bradycardia, hypertension, and irregular respirations.20 However, declining heart rates within normal limits (as opposed to frank bradycardia) up to 24 hours prior to clinical deterioration inferred by DHC are not currently recognized as a potential marker for neurologic deterioration.
Our observation that many trajectories were stable may be partially explained by modifying treatments including antihypertensives and insulin given by clinical providers. The absence of an increasing trajectory of glucose, contrary to prior studies that have found that baseline elevated glucose associated with increased edema,3,17 may lead investigators to consider whether increasing insulin requirements are a potential clinically relevant biomarker. We did not have access to insulin doses in this cohort. Other possible reasons for stable biomarker values may be because MLS \(\ge\)5mm does not sufficiently compress the brainstem to result in changes in serum laboratory values or vital signs.
The backward looking and estimation approach can augment simpler methods that consolidate repeatedly measured data using means or medians over time. Assessing how a biomarker is related to an outcome temporally is important to gain an understanding of phenotypic manifestations of disease progression. The approach can reveal patterns of expression level changes leading up to the event of interest in a way that lessens bias by accounting for repeatedly measured data of censored subjects who did not develop the event and truncating the use of data occurring after the event.
Moreover, our review and constructed tool of longitudinal trajectories over time in single individuals using the R Shiny App enables investigators to examine cases for hypothesis-generating purposes. Our approach of reviewing both forward- and backward-looking data provides a way to preliminarily test hypotheses seen in individual patients.
While variable trends can be incorporated into existing electronic medical record software including EPIC, they often exist as a single longitudinal trend, without adjustment for other factors that may either increase or decrease a patient’s risk of developing complications. Unlike the individualized patient charts we propose, variable trends also lack important intermediate milestones (including radiographic or clinical events) that can better alert clinicians to subtle or signature patterns prior to devastating late-stage events like herniation or neurologic deterioration.
Moreover, it is notable that the median trajectory patterns we observed were subtle, and did not include extremely abnormal values prior to any of our outcomes. We posit that subtle trajectories such as a decrease in beats per minute that remain within normal limits prior to DHC precede late occurring frank abnormalities. If validated, recognition of such trajectories could provide clinicians with the opportunity to obtain confirmatory diagnostic tests and intervene earlier before irreversible secondary catastrophic events.
Finally, as computational and visual tools become increasingly sophisticated, user-friendly, and widely available, we are approaching a time in which not systematically using real- or near-real time data to make management decisions will become unacceptable. Foundational work on clinically relevant trajectories will enable the development and validation of dynamic models that update over time with new information, providing more precision based quality care to individuals rather than population averages.
Our study has limitations. Our use of retrospective, observational data may be subject to residual confounding. Moreover, the backward looking modeling requires equally spaced data points. To fulfill this assumption, we assumed latent and linear progression between time points. Progression could theoretically be non-linear or have a different functional form on each time interval. Our cohort comes from two hospitals from a single region in which most of the population is White, and generalizability is uncertain. However, to amass a sufficiently large sample size of patients to better examine risk factors and outcomes, the large inclusion time was necessary. Moreover, the backward-looking estimation is a visualization tool and the shape of the trajectory is prone to subjectivity. However, rigorous methods of reviewing such trajectories must begin at the qualitative stage and are important for hypothesis generation for future, more definitive quantitative study. Finally, our observation of the outcome is dependent on diagnostic imaging, leading to potential nonrandom misclassification (e.g., elevations in WBC may occur after increasing MLS but appears prior due to when imaging was observed). However, given that we have consistent and increasing findings for three different outcomes, we are reassured of the potential veracity of the signal. Future directions include comparison of models that use dynamic predictors to current standards using static information only. Our patients may have been censored by death that occurred prior to outcome due to lack of imaging or withdrawal of life sustaining measures. Patients could be censored by DHC that occurred prior to radiographic outcomes.
Despite the limitations, we report on the largest dataset of MCA stroke patients with rigorously curated longitudinal laboratory, vital sign, and radiographic information of which we are aware. As electronically available repeatedly measured data are increasingly used, it becomes more important to develop robust methods to identify clinically relevant patterns that precede neurologic or systemic deterioration. Our methodologies serve as foundational hypothesis-generating work that can be applied to a variety of life-threatening conditions in the intensive care unit and our univariable findings will be used to inform the development of multivariable dynamic risk prediction models for life-threatening space-occupying mass effect.