Identifying critical thresholds
In this study, we performed a thorough thresholding analysis in hopes of identifying cerebrovascular reactivity thresholds that provide the most utility in deriving iICP. Beginning with PRx, a distinct threshold emerged during the Chi-square outcome analysis, + 0.05. This was the only threshold able to produce a statistically significant Chi-square for both mortality and favorable outcome prediction. Though it was unable to produce the strongest Spearman rank correlations with the cerebral physiologic insult burden measures, it was able to produce relatively strong correlations, and no other threshold was able to consistently produce the strongest correlations either. Additionally, the iICP derivation yield for + 0.05 was relatively high at 61.37%, only about 12% less than the largest yield observed. Based on the above findings, we identified + 0.05 as the likely optimal threshold for PRx-based iICP derivation. However, it is worth noting, that patients spent a relatively limited amount of time with severely elevated ICP, as shown in Table 1. This may have influenced the identification of + 0.05 as the optimal PRx threshold, as most of the data points are likely concentrated at lower ICP values and, by extension (given the relationship between ICP and PRx), at lower PRx values. Thus, the statistical strength observed at this threshold may, at least in part, reflect the underlying distribution of the dataset rather than a definitive physiologic inflection point.
For PAx-based iICP, the Chi-square analysis suggested a threshold of −0.20, as it was the threshold that produced the greatest peak when plotted. Furthermore, it was the only threshold to produce statistically significant Chi-squares when filtering for iICP with tight confidence bands, iICP.ci. However, this threshold falls within the range generally considered to represent intact reactivity (−1 to ~ 0), and is therefore, as discussed in the methods, unlikely to be a physiologically relevant threshold for iICP derivation. Although another peak was observed at + 0.55 on the mortality prediction plot, with a Chi-square only marginally smaller than that of − 0.20, this threshold was unable to produce a statistically significant Chi-square for favorable outcome prediction. It also failed to achieve statistical significance when iICP was filtered for small confidence bands, iICP.ci. Upon Spearman rank correlation testing, − 0.20 did produce the strongest correlation with any of the insult burden measures and even failed to produce a statistically significant correlation with both % time with CPP < 60mmHg and % time with CPP > 70mmHg. On the other hand, + 0.55 was able to produce the strongest correlations for these two cerebral physiologic insult burden metrics, as well as produce relatively strong associations for both % time with PRx > 0.25 and % time with RAC > 0. This seems promising for the threshold; however, + 0.55 is quite high considering that the highest identified critical PAx threshold for outcome prediction has been + 0.25 [36, 38]. Additionally, the iICP derivation yield associated with + 0.55 was abysmal at a mere 26.03%. Therefore, there appears to be no clear ideal threshold for PAx-based iICP derivation.
The Chi-square analysis for RAC-based iICP seemed to suggest − 0.45 as the ideal threshold, since it was the threshold that produced the most distinct Chi-square peak for both mortality and favorable outcome prediction. However, upon Spearman rank correlation testing, the threshold was unable to produce statistically significant correlations with % time with CPP > 70mmHg, % time with PRx > 0.25, and % time with PbtO2 < 20mmHg. Furthermore, this threshold is well within the negative range of RAC values, and thus, unlikely represents a physiologically relevant threshold for deriving iICP. Although + 0.30 was able to produce the strongest correlations for % time with CPP < 60 mmHg, % time with CPP > 70mmHg, and % time with PRx > 0.25, it was unable to produce any statistically significant associations with outcome prediction upon Chi-square analysis. Additionally, + 0.30 was associated with an incredibly low derivation yield (16.71%), making it unviable for iICP derivation. Due to the lack of any meaningful results, no ideal threshold for RAC-based iICP could be identified. Moreover, the utility of RAC for deriving iICP remains highly unclear in general, given the complexity of interpreting this index (RAC provides insight into not only cerebrovascular reactivity but also cerebral compensatory reserve) [20]. Therefore, more work is needed to evaluate the role of RAC in deriving personalized physiologic metrics.
This study provides the first comprehensive comparison of cerebrovascular reactivity thresholds for deriving iICP. Unlike previous studies that derived iICP using an arbitrarily selected threshold, we provided an in-depth evaluation of how threshold choice influences the performance of iICP. This work will inform future works, specifically algorithm development, in threshold selection. However, it is important to clarify that the thresholds identified in this study do not represent cerebrovascular reactivity cut-off values that are themselves most predictive of outcome. Rather, they solely reflect the optimal thresholds for deriving iICP, defined as those that produced iICP values most strongly associated with clinical outcomes and multimodal cerebral physiology.
Additional findings
Through this thresholding analysis, we made multiple additional interesting observations that deserve highlighting. Firstly, the findings of this study did not completely fall in line with the critical outcome prediction thresholds identified for the three cerebrovascular reactivity indices in recent literature. This is especially true for PAx and RAC, as we were unable to identify ideal thresholds for these. The current literature suggests critical thresholds in the ranges of 0 to + 0.25 and − 0.10 to + 0.05, for PAx and RAC, respectively [36, 38]. These critical thresholds were identified through association work between the cerebrovascular reactivity indices, as stand-alone parameters (not in deriving iICP), and long-term outcome using similar chi-square analyses. It would have been reasonable to expect that these critical thresholds would have been identified as the ideal thresholds for deriving iICP as well; however, this does not seem to be the case. For PRx, the literature has generally pointed towards a critical threshold within the range of + 0.25 to + 0.35 for mortality prediction [31, 36, 38]. However, there is some literature supporting a threshold of + 0.05. In one study by Sorrentino and colleagues, + 0.05 was identified as a critical threshold for favorable outcome prediction [31]. Furthermore, there is extensive pre-clinical animal literature suggesting that a PRx around 0 detects the lower limit of autoregulation [32, 40,41,42]. These studies provide some reassurance for the PRx threshold we identified here.
Second, PAx and RAC were associated with lower iICP derivation yields when compared to PRx. This mirrors recent findings from studies comparing the three indices for CPPopt derivation [43, 44]. One possible explanation for this is that, due to the highly controlled nature of ICP in the ICU setting, there may be too little variation in AMP to produce the needed variability in these indices to generate well fitted LOESS. This may result in identification of fewer iICP values (lower yield) or identification of inaccurate iICP values, both of which can blunt the ability of iICP to predict outcome. Therefore, it is likely that PRx represents the most practical cerebrovascular reactivity index for deriving iICP, since a low yields would significantly limit any clinical utility of iICP. However, we cannot make any conclusive statements on the underlying reasoning for this difference in yields. Additionally, it is important to note that these yields were produced using the entire recording periods of patients, and that a continuous multi-window weighted approach to iICP derivation, which would be necessary for clinical application, may produce different results. We, therefore, suggest that future iICP work not exclude these indices until further work has confirmed that they are inferior to PRx for iICP derivation.
Next, during Chi-square analysis, it was observed that favorable outcome prediction tended to produce greater Chi-squares than mortality prediction for PRx-based iICP, while producing smaller values than mortality prediction for RAC-based iICP. This suggests that PRx-based iICP is better at predicting favorable outcome than predicting mortality, while RAC-based iICP is better at predicting mortality than predicting favorable outcome. Also, for mortality prediction, RAC-based iICP produced greater Chi-squares than PAx-based iICP, which produced greater Chi-squares than PRx-based iICP. On the other hand, for favorable outcome prediction, PRx- and PAx-based iICP produced greater Chi-squares than RAC-based iICP. This suggests that RAC-based iICP may be best able to predict mortality, but the worst for predicting favorable outcome.
Regarding iICP.ci, it is interesting to see that at higher thresholds, more iICP values were filtered out than at the lower thresholds (see Supplemental Appendices A-C). This suggests that at higher thresholds, confidence in the accuracy of the identified iICP diminishes. iICP.ci also generally produced greater Chi-square values for outcome prediction, but lower yields, than compared to unfiltered iICP. This suggests that iICP based on confidence band size may result in greater ability to predict outcome, but at the expense of yield. Lastly, the subgroup analysis for age and sex was unable to identify any thresholds that were able to achieve significance for the ≥ 40 age group and female group. This may potentially suggest that various patient-specific factors can affect the utility of iICP, as well as the ideal cerebrovascular threshold for its derivation. However, this finding may be a result of differences in group sizes. Further work will be needed to investigate the role that patient demographics, injury severity, and treatment regimen has on iICP derivation.
Limitations
Despite the important findings uncovered in this thresholding analysis, there are a couple noteworthy limitations that must be addressed. Firstly, the main limitation of this thresholding analysis is that we generated iICP using patients’ entire recording periods. Therefore, it remains unknown whether the ideal thresholds identified here would be applicable to a continuously derived iICP. Currently, no continuous iICP algorithm exists; however, once one is developed, a further thresholding analysis may be necessary to confirm the idealness of the identified thresholds for deriving iICP in real-time.
Another limitation of this study is that the chi-square analysis failed to produce smooth plots where the chi-square values gradually increase, peak at an “ideal” threshold, and then gradually decrease (similar to what is seen for the yield curves). Rather, the generated plots present an erratic curve with sudden spikes. This questions whether the thresholds found to produce the strongest chi-square values are physiologically significant and not just mere statistical anomalies. Future work using datasets from outside the CAHR-TBI collaborative, such as high-resolution datasets from the CENTER-TBI and TRACK-TBI studies [45, 46], will be needed to validate our findings.
Since cerebral hemodynamics exhibit significant variation throughout the different phases of post-TBI recovery, it is highly possible that the ideal thresholds for iICP derivation vary over the course of a patient’s time in the ICU [36, 47]. In this study we did not consider such variations in cerebral physiology, thus limiting our findings. Future studies should consider stratifying monitoring periods across patients’ times in the ICU to better understand how these variations in cerebral physiology may affect optimal iICP derivation.Next, though the patient cohort used in this study was quite large, only a portion of them had PbtO2 recordings available (n = 106). This may have underpowered any tests involving this physiologic variable and may possibly explain why only one index-threshold pair was able to generate an iICP that produced a statistically significant association with % time with PbtO2 < 20mmHg. Future work with larger PbtO2 datasets is warranted to better shed light on the association between iICP and this important cerebral physiologic parameter.
Lastly, this study is limited by the scope of data available in the CAHR-TBI database. For instance, the database does not document whether patients underwent decompressive craniectomy, a procedure that recent literature suggests may influence cerebrovascular reactivity [48]. The absence of this information restricts our ability to account for a potentially important confounding variable. Furthermore, the lack of contemporary CT scoring systems, such as the Rotterdam or Helsinki CT scores, and the use of GOS, rather than the more detailed extended version (GOSE), limit our analyses and may affect the precision and generalizability of our findings.
Future directions
Unlike traditional static, population-based ICP thresholds, patient-specific ICP thresholds account for an individual’s dynamic cerebral autoregulatory status and may, in the future, enable treatment that is tailored to the individual’s specific physiologic needs. However, despite the promising preliminary findings regarding iICP, limited literature exists on the concept as of now. Additionally, the current state of the iICP concept is not conducive to clinical application. Firstly, the current algorithm requires a patient’s entire recording period, allowing only for the calculation of an “after the fact” threshold that is not usable to guide treatment. Moreover, it is only able to produce a singular threshold for a dataset and, therefore, does not take into account the dynamic nature of cerebral physiology over a patient’s time in the ICU. Another limitation of the current algorithm is that it fails to provide any assessment of curve fit characteristics. This prevents the clinical end-user from being able to gauge the quality of the output iICP value.
To circumvent these shortcomings, an algorithm that can continuously derive iICP in real-time is needed. Such an algorithm would require a sophisticated sliding multi-window weighted approach that, for each update interval (i.e. every minute), generates LOESS plots for various window lengths, scores plots based on a variety of factors (curve shape, confidence bands, data range, etc.), and calculates a weighted average to identify an iICP value. A similar strategy has been successfully leveraged in recent renditions of CPPopt [35, 49]. The algorithm should also present a summary of curve fit characteristics with each iICP calculation to allow for output quality assessment. Additionally, once a continuous algorithm is created, an assessment of whether the ideal CVR threshold for iICP derivation varies over different phases of the ICU stay (e.g., first 24 h vs. later periods) will be needed. This will provide valuable insight into how cerebrovascular reactivity impairment and iICP behavior evolve over time.
Following the development of such a continuous iICP derivation algorithm, thorough outcome analyses will be needed to provide preliminary insight into the prognostic utility of continuously derived iICP. Additionally, evaluation of the association between iICP and measures of cerebral physiologic insult burden will also be needed to shed light on the potential impact that iICP-directed care could have on minimizing secondary brain injury. However, to conclusively determine if iICP-directed care offers any real clinical benefit, a clinical trial would be needed.
Next, if iICP is to ever become implemented clinically, work will be needed to enable bedside implementation. This will require tailoring any continuously updating algorithm to the specific needs of the bedside environment and developing a user-interface that allows clinical end-users to efficiently use and adjust output values. Finally, while iICP represents a promising individualized approach to managing ICP, the integration of additional personalized cerebral physiologic metrics may further enhance the precision and utility of this tool. These include CPPopt, the mean arterial pressure optimum (MAPopt), and the bispectral index optimum (BISopt). In conjunction, these personalized metrics may help mitigate each other’s limitations, supporting a more comprehensive and effective strategy for bedside decision-making in neurocritical care.
