Temporal patterns of organ dysfunction after severe trauma

Organ dysfunction (OD) is a major threat to critically injured patients surviving through the resuscitation phase. Organ dysfunction composes a significant clinical challenge in the intensive care unit (ICU)-treated trauma patient and is associated with prolonged length of stay (LOS), adverse outcomes, and demanding resource utilisation. With more patients surviving through the early trauma phase, due to advances in trauma care systems, OD becomes an increasing determinant of outcome. There is a known association between admission variables such as injury severity, shock, age and comorbidity, and subsequent OD [1‐3]. However, the trajectories of OD, in terms of severity and composite temporal patterns and resolution, are complex and diverse. This aspect of trauma research has recently been identified as poorly investigated with a need for further studies [4].

Monitoring and treating in relation to temporal patterns of composite physiological states is inherent to ICU strategy. Each patient may be unique in their response. However, we expect human pathophysiology at some level to be groupable to distinct temporal patterns that are identifiable and reproducible. Understanding these patterns may help in timing of treatment strategies, pre-emptive identification of impending clinical decline [5‐7], and to identify cohorts that in the future may require precision treatment approaches identified with predictive enrichment strategies. Despite these discernible prerequisites and notable pursuits, ICU and trauma research have to date been more focused on temporally static approaches to pattern recognition. As such, important steps towards identifying composite phenotypes of disease in the ICU have been taken for ICU-related diseases such as ARDS [8] and sepsis [9], but rely predominantly on early presentation data and patterns, and take less into account information content found in the temporal evolution of disease. Thus, there is a clear need to align pattern recognition with clinical temporal reasoning and human-interpretable models, even if this necessitates complex methods of time-series analysis and pattern recognition.

Group-based trajectory modelling (GBTM), as developed and described by Nagin and Odgers [10], generally fit polynomials, over time for a group of trajectories, using maximum likelihood estimation to a predefined number of categories. It is a specialised application of the finite mixture model, a probabilistic modelling to identify subpopulations. The optimal number of trajectories may to some extent be objectified using Akaike or Bayesian information criterion (AIC or BIC), with BIC driving more parsimonious solutions. It has been used for a number of real-world clinical data analyses to identify temporal trajectories [11‐13] and affords a compelling method to reduce high-dimensional temporal data to a finite number of similar “type” trajectories.

The aim of this study was to analyse the temporal patterns of OD in ICU-treated critically injured patients. By using GBTM we attempt to identify and group sub-populations following a similar trajectory and characterise the temporal patterns of these groups in relation to both admission data and outcomes. In addition, we analyse the time from admission to trajectory group stabilisation in the identified sub-populations, as to discern if some subpopulations are less temporally predetermined and therefore potentially susceptible to treatment strategies during the ICU stay.

Six-hundred and sixty patients were included in the final model (see Fig. 1).

This is to our knowledge the first study using temporal clustering based on specific organ systems after severe trauma. We identified five groups with contrasting organ dysfunction trajectories, admission characteristics, and outcomes. The time to stabilisation into the final trajectory group differed notably between groups. This reflects the complexity of predicting the clinical course for some subsets of critically injured patients, but also provide opportunities to identify subgroups that may have avoidable deterioration or could benefit from specific treatments.

Post-injury OD is a well-recognised complication following severe trauma. This challenge is expected to increase with improved initial survival in modern trauma care systems, an ageing population, and rising burdens of comorbidity. The aetiology is not fully elucidated. There is a well-known association between injury severity, shock, and red blood cell transfusion [19, 20]. A factor advocated as a trigger of organ dysfunction is damage-associated molecular patterns (DAMPs) [21]. Substances like mitochondrial DNA, histones, and HMGB1 normally contained in the intra-cellular compartment released during tissue injury and hypoperfusion triggering an inflammatory response evoked by the innate immune system [22, 23]. There is an association between levels of circulating DAMPs and subsequent OD following trauma [22]. In addition, events occurring after trauma, such as infections and sepsis, may contribute to further OD and complicate prognostication of the clinical course.

Previous studies on post-injury OD are diverse. The use of different scoring systems in the assessment of OD complicates comparison between studies [24, 25]. There is also controversy as to the current trending of OD incidence, where some studies report that OD is a declining problem [26] and others report the contrary [27, 28]. In aggregate, in critically ill trauma patients a substantial unexplained heterogeneity of clinical course has been identified and advocated as a specific research challenge, calling for more exhaustive analyses [4, 29].

Few studies have addressed the pattern of OD over time, apart from onset times and impact on outcomes. Two previous studies analysing temporal patterns of OD did not assess specific organ systems but used scores of total OD [30, 31]. Thus, comparing the current study to previous work is problematic. Some resemblances are, however, noteworthy. One study of 440 UK patients, based on hierarchal clustering of total SOFA scores, identified three groups with contrasting trajectories, one dominated by TBI patients, one with severe injury and subsequently protracted OD and poor outcomes, and one with less injury, OD, and better outcomes [30]. In the current study we wanted to analyse both the temporal and organ-specific patterns of OD in an objective and discerning manner. The GBTM method has been shown to compare well with other methods of longitudinal models [32]. We used GBTM based on SOFA score for individual organs over time in order to identify groups with similar patterns. This method enabled us to use the full range of the temporal data by unsupervised differentiation of patients into latent trajectories, based on the joint inter-related development of their organ-based outcome patterns. Thus, this method allows to identify group assignment based on similarities in temporal combinations of specific patterns for each organ system, as opposed to previous studies using total OD without assessing specific organ systems. However, clustering of heterogeneous individual trajectories into fewer summarising trajectories is a task of dimensionality reduction and will always require some extent of simplification and approximation. With the GBTM method these trajectories were not assumed a priori as in subjective, often expert-based, categorisation of patients, but based on the true temporal profiles of the patients. Additionally, the method provides a measure of individual and group probability and, thus, certainty of trajectory group assignment.

This method generated five groups with different temporal patterns and outcomes. Group 1, the mild OD group had a favourable outcome with early resolution of OD and low mortality. Within this group there were, however, patients with diverging admission characteristics with some patients having an ISS of 50 and some being transfused with > 20 PRBCs units during the first 24 h. This cohort had the lowest median age and found their group trajectory rapidly after admission, suggesting that this subset of trauma patients can be identified at an early stage. Group 2 and group 3, the moderate and severe OD groups, were fairly similar in their admission parameters and had by far the longest time to trajectory stabilisation. These patients were in the mid-range in terms of injury severity and initial OD load as compared with the total cohort. However, the incidence of sepsis differed notably between the two groups, occurring more often in the severe OD group. The onset typically occurred a few days after admission possibly influencing trajectories and delaying stabilisation. These differences in time seen for trajectories to stabilise appear to reflect the heterogeneity amongst severely injured trauma patients that is often recognised clinically, where some patients develop an unexpectedly complicated clinical course, whereas others, with similar admission characteristics, do not. Our study suggests that while certain subsets of trauma patients will enter a clinical course that may be fairly well predicted in the early phase (group 5, TBI with OD, and group 1, mild OD), others have an undecided course (group 2, moderate OD, and group 3, severe OD) that may potentially be positively or negatively influenced by post-resuscitation events or treatments. This could suggest a potential window of therapy in these subgroups. In our study, sepsis appears related to a higher OD severity in the late stabilisers, which leads us to hypothesise that infectious complications may be a driver. The important implication of this is that potentially avoidable complications, arising after ICU admission, could be a reason that a significant number of patients with seemingly similar initial trajectories experienced different later patterns of organ failure. Consequently, their trajectories will be harder to define at an early stage.

The group of highly injured patients with extreme OD are readily recognisable clinically and pose a challenging entity in the ICU. In this study, group 4, extreme OD, was small and characterised by high injury severity where chest injuries were seen in 77%. Almost half of the patients were in shock at admission and two-thirds received massive transfusion, and these figures are in striking contrast to all other groups. The subsequent OD was considerable and sustained, affecting all organ systems. The incidence of liver and renal dysfunction was significant and increased during several days after admission. This was not seen in any other groups. Arguably, the injury severity, dose of shock, and massive transfusion seen in this group would be expected to generate a considerable release of DAMPs possibly contributing to the prominent OD seen in the post-injury phase. As expected, this group had a high mortality, but the median time to death was over a week for non-survivors, indicating a complicated clinical course, as seen in the vast need of organ supportive therapy. This small group of patients required a massive use of ICU recourses. The implication of such a highly identifiable group that distinguishes itself fairly early, and where mortality is extremely high, is that inflammatory and coagulopathic effects of shock and massive transfusion may need also to be a highlighted focus area of research in future trauma patients.

Traumatic brain injury represents a specific group of ICU patients. In this study, group 5, TBI with OD, had the worst outcomes of all groups in terms of mortality. Apart from CNS dysfunction subsequent to significant head injury the load of cardiovascular and respiratory OD was significant. This could in part be a confounder explained by the use of vasopressors to titrate optimal cerebral perfusion pressure and the need of extended controlled ventilation, as well as an expected high rate of respiratory infections, inflating non-CNS organ failure scores [33]. Traumatic brain injury is reported to be the main cause of death after trauma, and a high mortality in this group is expected [34, 35]. The median time to death was fairly short with 3.5 days for non-survivors as compared to group 4, suggesting a considerable proportion of refractory severe TBI. This group of patients with an almost 90% incidence of AIS head > 3 found their trajectory rapidly, and over 80% of these patients were assigned to their trajectory group by day 2. Thus, the TBI cohort is, not unexpectedly, highly discernible and exhibits an early defined trajectory of SOFA organ scores.

The methodology used here to group temporal trajectories provides a means with which to identify and separate phenotypes of trauma-related OD. Our findings exemplify the possibility of using not only expert-based or baseline data for clustering of patients, but instead to utilise the full range of available temporal data. Future research should be focused on external validation of our findings and using temporal clustering in other heterogeneous groups of intensive care patients such as for example cardiac arrest, sepsis, or ARDS.

We performed longitudinal clustering of severely injured trauma patients using data from all affected organs. Our study strives to objectify the temporal patterns of OD after admission to the ICU after severe trauma. Five distinct trajectories of organ failure after trauma were identified. Notable differences in admission characteristics, clinical course, and outcomes between the five groups generated by the GBTM modelling were seen. We found and describe some potentially important insights that were suggested by the groupings and temporal trajectories. These findings indicate subsets of patients with an initial undefined clinical course that might benefit from targeted support and future research focus. The findings also underline the heterogeneous course after trauma and the challenges faced in prognosticating the clinical course of these patients.