Variation in electroencephalography and neuroimaging for children receiving extracorporeal membrane oxygenation

Extracorporeal membrane oxygenation (ECMO) is used to support children with severe heart or lung failure. Neurologic injury sustained during ECMO support is one of the most feared complications for clinicians and families [1]. Patients on ECMO support are at risk of brain injury secondary to hypoperfusion, thromboembolism, hemorrhage, and reperfusion. A recent prospective multicenter observational study reported an incidence of 4% and 16% for ischemic and hemorrhagic neurologic complications of children on ECMO, respectively [2]. Also, single-center studies have demonstrated seizures in 17–40% of ECMO patients, many of which were only detectable on electroencephalography (EEG) [3‐6].

Clinicians deploy several strategies to identify brain injury in children because of the high risk of neurologic injury. For infants with an open fontanelle, cranial ultrasonography has been widely used and has the benefit of being portable and without radiation exposure [7‐9]. Imaging such as computed tomography (CT) and magnetic resonance imaging (MRI) may be used when there is suspicion of brain injury, but CT includes radiation exposure, and MRI is not possible during ECMO support. EEG can be used to detect seizures, monitor sedation, and may detect new focal abnormalities in the presence of ischemic or hemorrhagic brain injury [10, 11]. Less common neuromonitoring techniques such as transcranial Doppler (TCD) have also been suggested to monitor neurologic injury in patients receiving ECMO support [12‐14].

We sought to characterize the trends and hospital variation in the utilization of EEG and neuroimaging during ECMO hospitalizations among US children's hospitals using the Pediatric Health Information System (PHIS). This database allows for the assessment of current practices across US tertiary children's hospitals.

This study was determined to be not regulated as human subjects research by the University of Michigan Institutional Review Board and all ethical standards were followed (HUM00212344, 1/28/22). Data for this study were obtained from the Pediatric Health Information System (PHIS), an administrative database that contains inpatient, emergency department, ambulatory surgery, and observation encounter-level data from not-for-profit, tertiary care pediatric hospitals in the USA [15]. These hospitals are affiliated with the Children’s Hospital Association (Lenexa, KS). Data quality and reliability are assured through a joint effort between the Children's Hospital Association and participating hospitals. For the purposes of external benchmarking, participating hospitals provide discharge/encounter data including demographics, diagnoses, and procedures. Nearly all of these hospitals also submit resource utilization data (e.g., pharmaceuticals, imaging, and laboratory) into PHIS. Data are de-identified at the time of data submission, and data are subjected to reliability and validity checks before being included in the database. For this study, we included data from the 47 hospitals that contributed full data to PHIS at the time of the data request.

We included ECMO hospitalizations with a discharge date from January 1st, 2016 through December 31st, 2021. Hospitalizations were identified through ICD-10 procedure codes for ECMO (Additional file 1: Table S1). To allow subgroup analyses, we identified neonatal hospitalizations and hospitalization that included cardiac surgery. We defined neonatal admissions as those with a patient age of 28 days or less at the time of the first ECMO charge. We identified hospitalizations that included cardiac surgery through Risk adjustment for congenital heart surgery-2 (RACHS-2) model for ICD-10(c). RACHS-2 is an empirically derived, risk adjustment model that has been validated in two separate administrative data sources and compared to locally held clinical registry data [16].

We identified the following modalities through clinical transaction classification (CTC) codes used by PHIS: EEG, cranial ultrasound, head CT, brain MRI, and transcranial Doppler. We secondarily assessed the rates of seizure and stroke diagnoses and prescription of antiseizure medication. We determined stroke and seizure diagnoses through the presence of diagnosis codes [17‐20]. We report whether a patient was prescribed an antiseizure medication using CTC codes, excluding medications most likely to be used for sedation including benzodiazepines (Additional file 1: Table S1). Complex chronic conditions were determined according to complex chronic conditions classification software v2 developed by Feudtner et al. [21].

For each patient who received ECMO, the presence and number of charges for EEG and each imaging modality was recorded. We reported the proportion of patients who received each modality during the ECMO hospitalization. We also noted the median and interquartile range (IQR) for the frequency of each modality by patient who received the study. We reported patient outcomes of ECMO duration, hospital length of stay, and hospital mortality. We determined ECMO duration through the days of consecutive ECMO charges. If there was one day between ECMO charges, it was counted as a consecutive ECMO run. Bivariate comparisons were conducted through Pearson’s Chi-squared for categorical variables and Wilcoxon rank-sum for continuous variables. We tested for trends across years using a nonparametric test for trend [22].

For MRI and EEG, we grouped hospitals into quintiles based on the utilization of each modality. We report the frequency of a stroke diagnosis for each MRI quintile. We report the frequency of seizure diagnosis and antiseizure medication use for each EEG quintile. To assess variability between hospitals, we created a multilevel logistic regression model where hospitalizations were nested within hospitals, with the outcome of whether or not a patient received the modality. We modeled PHIS hospital as a random effect. For TCD, EEG, and head CT, we adjusted for patient-level covariates of patient age, duration of ECMO support, discharge year, and receipt of cardiac surgery a priori. For MRI, we also adjusted for patient in-hospital mortality, to account for patients needing to survive to decannulation from ECMO support to receive an MRI. To describe center-level variation, we reported the intraclass correlation coefficient and plotted the predicted probability of receiving each modality by center in a caterpillar plot [23, 24]. Analyses were completed in Stata (16.1, StataCorp LLC, College Station, TX).

In this analysis of children supported by ECMO at US Children's Hospitals, we found that there is significant variation in the use of EEG and neuroimaging modalities across hospitals, with the exception of cranial ultrasound, which was nearly universally applied in eligible infants. This center-level variation is especially pronounced in the use of EEG, which increased from 52% in 2016 to 72% in 2021. Among high and low EEG utilizers, the differences in rates of antiseizure medication use (36% vs 40%) and seizure diagnosis (16% vs 20%) were negligible. Similarly, there was a negligible difference in stroke diagnosis (17% vs 15%) between hospitals that were high MRI utilizers versus low utilizers.

The risk of neurologic injury during ECMO support is well known; however, the modalities and practices different centers use to monitor for and diagnose neurologic injury has not previously been reported. This study has several strengths. The PHIS database allows for an evaluation of practices across 47 children’s hospitals, has checks for reliability and validity, and has previously been used in pediatric critical care and ECMO research [25, 26]. The rates of stroke and seizure diagnoses seen in this cohort are consistent with the rates described in clinical research studies [2‐4], supporting the accuracy of the diagnosis codes used.

There was significant variation among children’s hospitals in the application of EEG with increasing use observed during the study period. This increase in use may reflect increasing awareness of seizures in children supported by ECMO and practice change toward more proactive neuromonitoring [3, 6, 27, 28]. There did not seem to be a connection in our data between the rate of EEG use at a hospital and the diagnosis of seizures or use of antiseizure medication. EEG has utility beyond seizure diagnosis, and its possible centers with high EEG use are deploying EEG to monitor sedation, assess for evolving asymmetries or other patterns that suggest new brain injury, or monitor for other neurologic dysfunction [10]. We observed a median of three days of EEG charges; importantly, we were unable to determine if this was continuous EEG monitoring or intermittent monitoring. We additionally were unable to determine if any patients’ EEGs were processed to include amplitude-integrated EEG or other quantitative EEG measures.[29] Another important consideration in the application of EEG monitoring is the significant cost and resource burden associated with continuous monitoring [30, 31]. A better understanding of the risk factors and timeline for neurologic injury in patients supported by ECMO may enable more cost-effective monitoring strategies.

There were notable differences between the subgroups analyzed. Compared with older children, neonates were more likely to receive an MRI during the hospitalization (46% vs 29%) but less likely to receive a CT (20% vs 49%). ELSO guidelines recommend considering MRI in all neonates or children less than 2 years of age who receive ECMO support, because subtle neurologic deficits may not be as apparent in this cohort [32]. This difference in use may represent awareness or adherence to these guidelines, or general approach toward higher utility of MRI in smaller children or patients where the clinical exam is less reliable. Yet, the diagnosis of stroke was similar between the high- and low-MRI utilization hospitals, which does not support the idea that routine use of neuroimaging may uncover neurologic injury that is not clinically apparent during the acute hospital admission [33, 34].

Compared to non-cardiac surgery hospitalizations, cardiac surgery hospitalizations including ECMO support were more likely to receive a CT (44% vs 33%) but less likely to receive an MRI (31% vs 39%). In the Bleeding and Thrombosis on ECMO (BATE) study, a cardiac indication for support was associated with an increased risk of both hemorrhagic and thrombotic events [2] although notably subsequent analyses have not shown a relationship between specific cardiac diagnoses or procedures with bleeding risk [35]. These differences and the findings noted above highlight that the variation in EEG and neuroimaging on ECMO likely not only has inter-hospital variation but also significant intra-hospital variation as neonatologists, intensivists, cardiologists, and surgeons approach the monitoring and diagnosis of brain injury differently.

The use of bedside tests that can detect brain injury or conditions that place a patient at risk for brain injury is appealing. Prior studies suggest that near-infrared spectroscopy is being widely used for this purpose; however, its use was not captured in this dataset [36‐38]. TCD has also been described for this purpose during ECMO support [13, 14, 39]. In this study, its use was limited to a small number of centers, where these noninvasive tests fit into the comprehensive neuromonitoring of a patient supported by ECMO should be a focus of future studies.

This study has important limitations. First, the use of billing data allowed for the characterization of practices across hospitals, but with outcomes limited to diagnosis codes, we are unable to accurately describe the results of the diagnostic tests. Similarly, we are unable to determine with accuracy whether the neurologic conditions described were acquired during the hospitalization or were pre-existing. Characterizing neurologic morbidities is essential in designing protocols for monitoring and screening. Second, the PHIS dataset contains admissions from US children’s hospitals and may not be generalizable or representative of practices at non-children's hospitals or those outside the USA. Third, it is possible that increased neuromonitoring reflects the contribution of neurology consultants; the PHIS dataset does not provide information about use of neurology consultation services and whether centers have protocolized neuromonitoring or it is clinician-dependent.

Clinical teams have used many modalities to screen, diagnose, and monitor for neurologic injury in this high-risk population with minimal evidence to guide their application. This has resulted in significant variation among centers providing ECMO support and between patient populations. There is an opportunity for prospective studies and national organizations to identify patients at the highest risk and provide guidance on best practices for screening and monitoring. Importantly, the identification of injury is essential but insufficient in the comprehensive care of a patient with a brain injury. As we develop best practices for monitoring, we also need to support the multidisciplinary teams, including pediatric intensivists, neurologists, and rehabilitation specialists to develop plans to treat evolving injury and engage children in long-term neurodevelopmental support.