gms | German Medical Science

Background: Life-threatening conditions such as sepsis commonly affect multiple organ systems concurrently. Statistical network analysis of parameters representing the main organ systems has previously been shown to be potentially useful for the evaluation of the prognosis of critically ill patients in the intensive care unit (ICU). However, so far, the difference in organ system networks between a survivor and a non-survivor group has been evaluated at ICU admission only. The dynamics of networks and potential differences in network dynamics in patients surviving and not surviving the ICU stay has not yet been investigated. Here, we report preliminary findings from our longitudinal and cross-sectional comparison of organ system networks at two different, clinically relevant time points during the ICU stay: after admission and prior to death in non-survivors and a matching time point during ICU treatment in survivors.

Methods: We analyzed the interactions between representative variables of 9 different organ systems based on consecutive admissions to the surgical ICU of the University Medical Centre Mannheim, Germany.

We combined risk set sampling and propensity score matching to identify appropriate controls for the deceased patients from the electronic medical records. This way, controls were treated for at least as long as their matching case, with their index time point corresponding to their matching case's recorded time of death. Both had to have measurements for all network parameters at least once between 0-24 hours after admission (admission networks) and once during 48-24 hours before death or index time point (event networks). The propensity score was then derived from baseline parameters and scores, leaving 149 case-control pairs after matching. Admission and event networks were compared within and between cases and controls, leading to 4 comparisons.

Network structures were fitted using three conceptually different approaches: (1) Spearman correlations, (2) partial Spearman correlations, and (3) the EBICglasso regularization method. Community and cluster detection in the networks was performed via the Girvan-Newman algorithm. Differences between networks were further evaluated using multiple similarity indices, as well as permutation-based tests, taking account of the unpaired and paired settings in cross-sectional and longitudinal comparisons, respectively.

Results: The outcomes of our network comparisons depend on the network estimation approach.

Fitting the network via (partial) Spearman correlations showed that while the admission network structures are fairly similar among survivors and non-survivors, the corresponding event networks differ substantially. Particularly, the event network structure of non-survivors is severely disrupted with only few edges remaining, whereas that of the survivors has far more edges aggregated in few clusters.

A negative correlation of bilirubin concentration and platelet count is present in all networks.

While CRP typically has many edges in all 3 other networks, it shows no edge in the survivors' event network.

Conclusion: Our work suggests meaningful differences in the event networks between survivors and non-survivors. Longitudinal network comparison may reveal relevant changes of organ system interactions in critically ill patients in the course of ICU treatment.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.