Validation of a model to predict electroencephalographic seizures in critically ill children

Survey of Pediatric ICU EEG Monitoring-Reassessment After a Decade

PURPOSE

In 2011, the authors conducted a survey regarding continuous EEG (CEEG) utilization in critically ill children. In the interim decade, the literature has expanded, and guidelines and consensus statements have addressed CEEG utilization. Thus, the authors aimed to characterize current practice related to CEEG utilization in critically ill children.

METHODS

The authors conducted an online survey of pediatric neurologists from 50 US and 12 Canadian institutions in 2022.

RESULTS

The authors assessed responses from 48 of 62 (77%) surveyed institutions. Reported CEEG indications were consistent with consensus statement recommendations and included altered mental status after a seizure or status epilepticus, altered mental status of unknown etiology, or altered mental status with an acute primary neurological condition. Since the prior survey, there was a 3- to 4-fold increase in the number of patients undergoing CEEG per month and greater use of written pathways for ICU CEEG. However, variability in resources and workflow persisted, particularly regarding technologist availability, frequency of CEEG screening, communication approaches, and electrographic seizure management approaches.

CONCLUSIONS

Among the surveyed institutions, which included primarily large academic centers, CEEG use in pediatric intensive care units has increased with some practice standardization, but variability in resources and workflow were persistent.

Periodic Discharges in Critically Ill Children: Predictors and Outcome

OBJECTIVES

We aimed to identify clinical and EEG monitoring characteristics associated with generalized, lateralized, and bilateral-independent periodic discharges (GPDs, LPDs, and BIPDs) and to determine which patterns were associated with outcomes in critically ill children.

METHODS

We performed a prospective observational study of consecutive critically ill children undergoing continuous EEG monitoring, including standardized scoring of GPDs, LPDs, and BIPDs. We identified variables associated with GPDs, LPDs, and BIPDs and assessed whether each pattern was associated with hospital discharge outcomes including the Glasgow Outcome Scale-Extended Pediatric version (GOS-E-Peds), Pediatric Cerebral Performance Category (PCPC), and mortality.

RESULTS

PDs occurred in 7% (91/1,399) of subjects. Multivariable logistic regression indicated that patients with coma (odds ratio [OR], 3.45; 95% confidence interval [CI]: 1.55, 7.68) and abnormal EEG background category (OR, 6.85; 95% CI: 3.37, 13.94) were at increased risk for GPDs. GPDs were associated with mortality (OR, 3.34; 95% CI: 1.24, 9.02) but not unfavorable GOS-E-Peds (OR, 1.93; 95% CI: 0.88, 4.23) or PCPC (OR, 1.64; 95% CI: 0.75, 3.58). Patients with acute nonstructural encephalopathy did not experience LPDs, and LPDs were not associated with mortality or unfavorable outcomes. BIPDs were associated with mortality (OR, 3.68; 95% CI: 1.14, 11.92), unfavorable GOS-E-Peds (OR, 5.00; 95% CI: 1.39, 18.00), and unfavorable PCPC (OR, 5.96; 95% CI: 1.65, 21.46).

SIGNIFICANCE

Patients with coma or more abnormal EEG background category had an increased risk for GPDs and BIPDs, and no patients with an acute nonstructural encephalopathy experienced LPDs. GPDs were associated with mortality and BIPDs were associated with mortality and unfavorable outcomes, but LPDs were not associated with unfavorable outcomes.

EEG Monitoring in Critically Ill Children: Establishing High-Yield Subgroups

PURPOSE

Continuous EEG monitoring (CEEG) is increasingly used to identify electrographic seizures (ES) in critically ill children, but it is resource intense. We aimed to assess how patient stratification by known ES risk factors would impact CEEG utilization.

METHODS

This was a prospective observational study of critically ill children with encephalopathy who underwent CEEG. We calculated the average CEEG duration required to identify a patient with ES for the full cohort and subgroups stratified by known ES risk factors.

RESULTS

ES occurred in 345 of 1,399 patients (25%). For the full cohort, an average of 90 hours of CEEG would be required to identify 90% of patients with ES. If subgroups of patients were stratified by age, clinically evident seizures before CEEG initiation, and early EEG risk factors, then 20 to 1,046 hours of CEEG would be required to identify a patient with ES. Patients with clinically evident seizures before CEEG initiation and EEG risk factors present in the initial hour of CEEG required only 20 (<1 year) or 22 (≥1 year) hours of CEEG to identify a patient with ES. Conversely, patients with no clinically evident seizures before CEEG initiation and no EEG risk factors in the initial hour of CEEG required 405 (<1 year) or 1,046 (≥1 year) hours of CEEG to identify a patient with ES. Patients with clinically evident seizures before CEEG initiation or EEG risk factors in the initial hour of CEEG required 29 to 120 hours of CEEG to identify a patient with ES.

CONCLUSIONS

Stratifying patients by clinical and EEG risk factors could identify high- and low-yield subgroups for CEEG by considering ES incidence, the duration of CEEG required to identify ES, and subgroup size. This approach may be critical for optimizing CEEG resource allocation.

Late-Onset Findings During Extended EEG Monitoring Are Rare in Critically Ill Children

PURPOSE

Electrographic seizures (ES) are common in critically ill children undergoing continuous EEG (CEEG) monitoring, and previous studies have aimed to target limited CEEG resources to children at highest risk of ES. However, previous studies have relied on observational data in which the duration of CEEG was clinically determined. Thus, the incidence of late occurring ES is unknown. The authors aimed to assess the incidence of ES for 24 hours after discontinuation of clinically indicated CEEG.

METHODS

This was a single-center prospective study of nonconsecutive children with acute encephalopathy in the pediatric intensive care unit who underwent 24 hours of extended research EEG after the end of clinical CEEG. The authors assessed whether there were new findings that affected clinical management during the extended research EEG, including new-onset ES.

RESULTS

Sixty-three subjects underwent extended research EEG. The median duration of the extended research EEG was 24.3 hours (interquartile range 24.0-25.3). Three subjects (5%) had an EEG change during the extended research EEG that resulted in a change in clinical management, including an increase in ES frequency, differential diagnosis of an event, and new interictal epileptiform discharges. No subjects had new-onset ES during the extended research EEG.

CONCLUSIONS

No subjects experienced new-onset ES during the 24-hour extended research EEG period. This finding supports observational data that patients with late-onset ES are rare and suggests that ES prediction models derived from observational data are likely not substantially underrepresenting the incidence of late-onset ES after discontinuation of clinically indicated CEEG.

Optimizing EEG monitoring in critically ill children at risk for electroencephalographic seizures

OBJECTIVE

Strategies are needed to optimally deploy continuous EEG monitoring (CEEG) for electroencephalographic seizure (ES) identification and management due to resource limitations. We aimed to construct an efficient multi-stage prediction model guiding CEEG utilization to identify ES in critically ill children using clinical and EEG covariates.

METHODS

The largest prospective single-center cohort of 1399 consecutive children undergoing CEEG was analyzed. A four-stage model was developed and trained to predict whether a subject required additional CEEG at the conclusion of each stage given their risk of ES. Logistic regression, elastic net, random forest, and CatBoost served as candidate methods for each stage and were evaluated using cross validation. An optimal multi-stage model consisting of the top-performing stage-specific models was constructed.

RESULTS

When evaluated on a test set, the optimal multi-stage model achieved a cumulative specificity of 0.197 and cumulative F1 score of 0.326 while maintaining a high minimum cumulative sensitivity of 0.938. Overall, 11 % of test subjects with ES were removed from the model due to a predicted low risk of ES (falsely negative subjects). CEEG utilization would be reduced by 32 % and 47 % compared to performing 24 and 48 h of CEEG in all test subjects, respectively. We developed a web application called EEGLE (EEG Length Estimator) that enables straightforward implementation of the model.

CONCLUSIONS

Application of the optimal multi-stage ES prediction model could either reduce CEEG utilization for patients at lower risk of ES or promote CEEG resource reallocation to patients at higher risk for ES.

Association of EEG Background With Diffusion-Weighted Magnetic Resonance Neuroimaging and Short-Term Outcomes After Pediatric Cardiac Arrest

BACKGROUND AND OBJECTIVES

EEG and MRI features are independently associated with pediatric cardiac arrest (CA) outcomes, but it is unclear whether their combination improves outcome prediction. We aimed to assess the association of early EEG background category with MRI ischemia after pediatric CA and determine whether addition of MRI ischemia to EEG background features and clinical variables improves short-term outcome prediction.

METHODS

This was a single-center retrospective cohort study of pediatric CA with EEG initiated ≤24 hours and MRI obtained ≤7 days of return of spontaneous circulation. Initial EEG background was categorized as normal, slow/disorganized, discontinuous/burst-suppression, or attenuated-featureless. MRI ischemia was defined as percentage of brain tissue with apparent diffusion coefficient (ADC) <650 × 10-6 mm2/s and categorized as high (≥10%) or low (<10%). Outcomes were mortality and unfavorable neurologic outcome (Pediatric Cerebral Performance Category increase ≥1 from baseline resulting in ICU discharge score ≥3). The Kruskal-Wallis test evaluated the association of EEG with MRI. Area under the receiver operating characteristic (AUROC) curve evaluated predictive accuracy. Logistic regression and likelihood ratio tests assessed multivariable outcome prediction.

RESULTS

We evaluated 90 individuals. EEG background was normal in 16 (18%), slow/disorganized in 42 (47%), discontinuous/burst-suppressed in 12 (13%), and attenuated-featureless in 20 (22%) individuals. The median percentage of MRI ischemia was 5% (interquartile range 1-18); 32 (36%) individuals had high MRI ischemia burden. Twenty-eight (31%) individuals died, and 58 (64%) had unfavorable neurologic outcome. Worse EEG background category was associated with more MRI ischemia (p < 0.001). The combination of EEG background and MRI ischemia burden had higher predictive accuracy than EEG alone (AUROC: mortality: 0.92 vs 0.87, p = 0.03) or MRI alone (AUROC: mortality: 0.92 vs 0.84, p = 0.02; unfavorable: 0.83 vs 0.73, p < 0.01). Addition of percentage of MRI ischemia to clinical variables and EEG background category improved prediction for mortality (χ2 = 19.1, p < 0.001) and unfavorable neurologic outcome (χ2 = 4.8, p = 0.03) and achieved high predictive accuracy (AUROC: mortality: 0.97; unfavorable: 0.92).

DISCUSSION

Early EEG background category was associated with MRI ischemia after pediatric CA. Combining EEG and MRI data yielded higher outcome predictive accuracy than either modality alone. The addition of MRI ischemia to clinical variables and EEG background improved short-term outcome prediction.

Electrographic Seizure Characteristics and Electrographic Status Epilepticus Prediction

PURPOSE

We aimed to characterize electrographic seizures (ES) and electrographic status epilepticus (ESE) and determine whether a model predicting ESE exclusively could effectively guide continuous EEG monitoring (CEEG) utilization in critically ill children.

METHODS

This was a prospective observational study of consecutive critically ill children with encephalopathy who underwent CEEG. We used descriptive statistics to characterize ES and ESE, and we developed a model for ESE prediction.

RESULTS

ES occurred in 25% of 1,399 subjects. Among subjects with ES, 23% had ESE, including 37% with continuous seizures lasting >30 minutes and 63% with recurrent seizures totaling 30 minutes within a 1-hour epoch. The median onset of ES and ESE occurred 1.8 and 0.18 hours after CEEG initiation, respectively. The optimal model for ESE prediction yielded an area under the receiver operating characteristic curves of 0.81. A cutoff selected to emphasize sensitivity (91%) yielded specificity of 56%. Given the 6% ESE incidence, positive predictive value was 11% and negative predictive value was 99%. If the model were applied to our cohort, then 53% of patients would not undergo CEEG and 8% of patients experiencing ESE would not be identified.

CONCLUSIONS

ESE was common, but most patients with ESE had recurrent brief seizures rather than long individual seizures. A model predicting ESE might only slightly improve CEEG utilization over models aiming to identify patients at risk for ES but would fail to identify some patients with ESE. Models identifying ES might be more advantageous for preventing ES from evolving into ESE.

Review of Noninvasive Neuromonitoring Modalities in Children II: EEG, qEEG

Critically ill children with acute neurologic dysfunction are at risk for a variety of complications that can be detected by noninvasive bedside neuromonitoring. Continuous electroencephalography (cEEG) is the most widely available and utilized form of neuromonitoring in the pediatric intensive care unit. In this article, we review the role of cEEG and the emerging role of quantitative EEG (qEEG) in this patient population. cEEG has long been established as the gold standard for detecting seizures in critically ill children and assessing treatment response, and its role in background assessment and neuroprognostication after brain injury is also discussed. We explore the emerging utility of both cEEG and qEEG as biomarkers of degree of cerebral dysfunction after specific injuries and their ability to detect both neurologic deterioration and improvement.

Validation of a Model for Targeted EEG Monitoring Duration in Critically Ill Children

PURPOSE

Continuous EEG monitoring (CEEG) to identify electrographic seizures (ES) in critically ill children is resource intense. Targeted strategies could enhance implementation feasibility. We aimed to validate previously published findings regarding the optimal CEEG duration to identify ES in critically ill children.

METHODS

This was a prospective observational study of 1,399 consecutive critically ill children with encephalopathy. We validated the findings of a multistate survival model generated in a published cohort ( N = 719) in a new validation cohort ( N = 680). The model aimed to determine the CEEG duration at which there was <15%, <10%, <5%, or <2% risk of experiencing ES if CEEG were continued longer. The model included baseline clinical risk factors and emergent EEG risk factors.

RESULTS

A model aiming to determine the CEEG duration at which a patient had <10% risk of ES if CEEG were continued longer showed similar performance in the generation and validation cohorts. Patients without emergent EEG risk factors would undergo 7 hours of CEEG in both cohorts, whereas patients with emergent EEG risk factors would undergo 44 and 36 hours of CEEG in the generation and validation cohorts, respectively. The <10% risk of ES model would yield a 28% or 64% reduction in CEEG hours compared with guidelines recommending CEEG for 24 or 48 hours, respectively.

CONCLUSIONS

This model enables implementation of a data-driven strategy that targets CEEG duration based on readily available clinical and EEG variables. This approach could identify most critically ill children experiencing ES while optimizing CEEG use.

Periodic and rhythmic patterns in critically ill children: Incidence, interrater agreement, and seizures

OBJECTIVES

We aimed to determine the incidence of periodic and rhythmic patterns (PRP), assess the interrater agreement between electroencephalographers scoring PRP using standardized terminology, and analyze associations between PRP and electrographic seizures (ES) in critically ill children.

METHODS

This was a prospective observational study of consecutive critically ill children undergoing continuous electroencephalographic monitoring (CEEG). PRP were identified by one electroencephalographer, and then two pediatric electroencephalographers independently scored the first 1-h epoch that contained PRP using standardized terminology. We determined the incidence of PRPs, evaluated interrater agreement between electroencephalographers scoring PRP, and evaluated associations between PRP and ES.

RESULTS

One thousand three hundred ninety-nine patients underwent CEEG. ES occurred in 345 (25%) subjects. PRP, ES + PRP, and ictal-interictal continuum (IIC) patterns occurred in 142 (10%), 81 (6%), and 93 (7%) subjects, respectively. The most common PRP were generalized periodic discharges (GPD; 43, 30%), lateralized periodic discharges (LPD; 34, 24%), generalized rhythmic delta activity (GRDA; 34, 24%), bilateral independent periodic discharges (BIPD; 14, 10%), and lateralized rhythmic delta activity (LRDA; 11, 8%). ES risk varied by PRP type (p < .01). ES occurrence was associated with GPD (odds ratio [OR] = 6.35, p < .01), LPD (OR = 10.45, p < .01), BIPD (OR = 6.77, p < .01), and LRDA (OR = 6.58, p < .01). Some modifying features increased the risk of ES for each of those PRP. GRDA was not significantly associated with ES (OR = 1.34, p = .44). Each of the IIC patterns was associated with ES (OR = 6.83-8.81, p < .01). ES and PRP occurred within 6 h (before or after) in 45 (56%) subjects.

SIGNIFICANCE

PRP occurred in 10% of critically ill children who underwent CEEG. The most common patterns were GPD, LPD, GRDA, BIPD, and LRDA. The GPD, LPD, BIPD, LRDA, and IIC patterns were associated with ES. GRDA was not associated with ES.

Evaluation of Seizure Risk in Infants After Cardiopulmonary Bypass in the Absence of Deep Hypothermic Cardiac Arrest

Background

Guidelines recommend evaluation for electrographic seizures in neonates and children at risk, including after cardiopulmonary bypass (CPB). Although initial research using screening electroencephalograms (EEGs) in infants after CPB found a 21% seizure incidence, more recent work reports seizure incidences ranging 3–12%. Deep hypothermic cardiac arrest was associated with increased seizure risk in prior reports but is uncommon at our institution and less widely used in contemporary practice. This study seeks to establish the incidence of seizures among infants following CPB in the absence of deep hypothermic cardiac arrest and to identify additional risk factors for seizures via a prediction model.

Methods

A retrospective chart review was completed of all consecutive infants ≤ 3 months who received screening EEG following CPB at a single center within a 2-year period during 2017–2019. Clinical and laboratory data were collected from the perioperative period. A prediction model for seizure risk was fit using a random forest algorithm, and receiver operator characteristics were assessed to classify predictions. Fisher’s exact test and the logrank test were used to evaluate associations between clinical outcomes and EEG seizures.

Results

A total of 112 infants were included. Seizure incidence was 10.7%. Median time to first seizure was 28.1 h (interquartile range 18.9–32.2 h). The most important factors in predicting seizure risk from the random forest analysis included postoperative neuromuscular blockade, prematurity, delayed sternal closure, bypass time, and critical illness preoperatively. When variables captured during the EEG recording were included, abnormal postoperative neuroimaging and peak lactate were also highly predictive. Overall model accuracy was 90.2%; accounting for class imbalance, the model had excellent sensitivity and specificity (1.00 and 0.89, respectively).

Conclusions

Seizure incidence was similar to recent estimates even in the absence of deep hypothermic cardiac arrest. By employing random forest analysis, we were able to identify novel risk factors for postoperative seizure in this population and generate a robust model of seizure risk. Further work to validate our model in an external population is needed.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12028-021-01313-1.

Modulation in time of the interictal spiking pattern related to epileptic seizures

OBJECTIVE

To test the hypothesis that significant changes in the occurrence of interictal epileptiform electroencephalography (EEG) discharges (EDs) are associated with seizures: while some EDs are pro-convulsive, increasing at seizure-occurrence, others are protective, showing decrease related to seizures.

METHODS

We analyzed 102 consecutive, long-term video-EEG monitoring sessions, from 98 patients. Using a semi-automated spike-detection method, we quantified the occurrence of EDs, grouped according to their location and morphology (clusters) and we constructed graphical representation of data, showing changes in time of the spiking patterns (spike-histograms). We compared the spike-histograms with the time-points of the seizures (pre-, peri- and postictal changes).

RESULTS

Totally 179 ED-clusters were identified. Modulation of the spiking pattern, associated with seizures, was observed in 66 clusters (37%), from 47 patients (48%). Most of these changes (40 clusters; 61%) were related to increase in the spiking-pattern.

CONCLUSIONS

Changes in spiking-pattern were associated with more than one third of the EDs. Both increasing and decreasing patterns were observed.

SIGNIFICANCE

EDs are more often pro-convulsive, with increasing spiking patterns associated with seizures. However, in more than one third of the ED clusters modulated by seizures, the spiking pattern decreased, raising the possibility of an anticonvulsive function of these discharges.

Machine learning models to predict electroencephalographic seizures in critically ill children

OBJECTIVE

To determine whether machine learning techniques would enhance our ability to incorporate key variables into a parsimonious model with optimized prediction performance for electroencephalographic seizure (ES) prediction in critically ill children.

METHODS

We analyzed data from a prospective observational cohort study of 719 consecutive critically ill children with encephalopathy who underwent clinically-indicated continuous EEG monitoring (CEEG). We implemented and compared three state-of-the-art machine learning methods for ES prediction: (1) random forest; (2) Least Absolute Shrinkage and Selection Operator (LASSO); and (3) Deep Learning Important FeaTures (DeepLIFT). We developed a ranking algorithm based on the relative importance of each variable derived from the machine learning methods.

RESULTS

Based on our ranking algorithm, the top five variables for ES prediction were: (1) epileptiform discharges in the initial 30 minutes, (2) clinical seizures prior to CEEG initiation, (3) sex, (4) age dichotomized at 1 year, and (5) epileptic encephalopathy. Compared to the stepwise selection-based approach in logistic regression, the top variables selected by our ranking algorithm were more informative as models utilizing the top variables achieved better prediction performance evaluated by prediction accuracy, AUROC and F1 score. Adding additional variables did not improve and sometimes worsened model performance.

CONCLUSION

The ranking algorithm was helpful in deriving a parsimonious model for ES prediction with optimal performance. However, application of state-of-the-art machine learning models did not substantially improve model performance compared to prior logistic regression models. Thus, to further improve the ES prediction, we may need to collect more samples and variables that provide additional information.