The primary purpose of this pilot study was to determine whether novel physiomarkers for eICP events can be detected using machine learning methods in the physiologic data streams from continuous bedside monitors of children with severe brain injury. Secondarily, we benchmarked the predictive performance of the modeling approaches we used.
Though we expected features derived from the heart rate signal to have high differential expression because of the often-observed sympathetic surges which precede herniation events, we did not find this to be the case (Fig. 5). One reason may be because our analysis employed numerical HR series at minute-to-minute resolution and statistical feature surrogates of heart rate variability, whereas traditional heart rate variability analyses rely upon R-R interval determinations from EKG waveforms and reflect changes in autonomic state [20]. Additionally, the inherent artifact burden in the HR signal – which we did not filter – may have reduced the contributions of this signal.
We found, instead, that features contained within the blood pressure signal were consistently the most differentially expressed. This suggests that (a) physiomarkers for eICP events are concentrated within the blood pressure signal, and (b) they are accessible to machine learning analyses. The finding is consistent with the clinical observation that hemodynamic changes occur prior to eICP events, although the typical clinically observable changes in vital signs are usually coarse and very proximate to events. The strong contribution of blood pressure features to our XGB predictive model not only endured across every time window we tested, but they progressively overshadowed ICP-derived feature importance earlier in the 4-hour records (Fig. 5). This was true in the context of a relatively sustained model performance across this time period (see AUPRC values). Though predictions 30 minutes prior to eICP events would enable most early interventions, the possibility of obtaining such insight hours prior to events is provocative. The notable implications of these results are that such physiomarkers could be used to predict elevated ICP in the presence of monitoring, and they may also have the potential to discriminate which children would benefit from placement of an ICP monitor.
We do not believe this result can be causatively explained by clinical differences among cases and controls. Children of different ages can have significantly disparate hemodynamic parameters, but the ages of case and control records were not different. Although 20% more control records were derived from patients who were not on either vasoactive or anti-hypertensive infusions, the mean VIS between cases and controls was similar. That said, we cannot predict how vasoactive/inotropic agents, antihypertensives, or hemodynamic-influencing sedatives impact these physiomarkers. Also, despite more control records coming from patients with craniectomies, this surgery would only serve to reduce the chances of having eICP events and not, in itself, impact hemodynamics.
The limited contribution of MAP from among the hemodynamic signals stands in interesting contrast to the findings of Güiza et al., who also reported on a machine learning model to predict eICP events using ICP and MAP as model inputs. Using 178 patients and 982 eICP events for model development, their model predicted eICP events 30 minutes prior to occurrence with an AUROC of 0.87 in an initial study [7] and of 0.90 when externally validated in adults [8]. With a pediatric external validation cohort, these authors found a lower AUROC of 0.79, comparable to our benchmark AUROC of 0.82. This may have been because their development cohort contained only 13 pediatric patients. Sensitivities were similar (92% and 93%), though our model achieved better specificity (48% vs 65%). Güiza et al.’s excellent predictive performance using MAP and ICP alone may be a reflection of their large training cohorts as well as their predominantly adult population which has less vital sign variation than children. It is also possible that different machine learning modeling approaches – and importantly, the features which are examined – may modulate the measured contribution of each physiologic signal. Nevertheless, our findings that differentially expressed physiomarkers are concentrated within the blood pressure signal as well as the ICP signal are consistent with Güiza et al.’s modeling results.
Though our exclusively pediatric cohort of different ages and mixed brain injury pathologies who underwent standard brain injury management may make the results of our small pilot study generalizable, they require externally validation in a larger cohort. There are several limitations. Our sample size was small for machine learning. Though minute-to-minute data resolution may be accessible to more ICUs, it may have also constrained our model performance. Measures of change and variability were limited in our feature selection, and our design did not consider changes in feature values over time across analysis windows. Our records were curated and imbalanced. We did not study the impact of vasoactive agents or sedatives (e.g. propofol, dexmedetomidine) which may have hemodynamic effects, or of management conditions (e.g. craniectomy, type of ICP monitor, etc.), though their use appear similar across cases and controls.
In summary, we find that discriminatory physiomarkers discernable by machine learning methods are concentrated within blood pressure data for up to 4 hours prior to future eICP events in children with severe brain injuries, and models using them demonstrate strong benchmark performance. Taken in aggregate, our findings support the idea that future attempts at high fidelity predictive modeling of eICP events should leverage features contained within hemodynamic signals.