Diffuse cerebral edema is a well-known complication of hypoxic-ischemic brain injury,¹ affecting 16–30%² of post-cardiac arrest patients. It is the most frequently identified abnormality on early head computed tomography (CT)^3,4 and is thought to be a major contributor to morbidity, given its association with poor neurologic and survival outcomes.^4–6 Diffuse cerebral edema arises from disrupted ionic and osmotic balance following hypoxic-ischemic injury, leading to water accumulation and cellular swelling in the brain.^7,8 Managing diffuse cerebral edema after cardiac arrest remains challenging because limited data exists to guide targeted treatments for elevated intracranial pressure in this context. As a result, clinicians rely on standard ICP-lowering therapies, yet edema is often recognized only after clinical or radiographic deterioration, limiting opportunities to trial or develop proactive interventions.^5,9–13 Diffuse cerebral edema is typically identified after alarming clinical signs such as fixed, dilated pupils,¹⁴ and on imaging by the widespread loss of gray-white differentiation, sulcal effacement, compression of the basal cisterns and ventricles, culminating in transtentorial herniation.^6,15 However, in its earlier stages, evolving cerebral edema may go unrecognized due to inconsistent imaging intervals¹⁵ and clinical exam limitations in comatose patients due to both injury severity and confounding effects of sedation,¹⁶ temperature control,¹⁵ as well as metabolic derangements.¹⁷

Electroencephalography (EEG) is non-invasive and widely recommended for seizure detection and prognostication in post-cardiac arrest patients.^18–20 Diffuse cerebral edema on CT has been associated with an overall suppressed EEG background,^18,21,22 as well as changes in quantitative EEG features,^16,22 suggesting that EEG waveform characteristics can be leveraged to automatically detect cerebral edema development at earlier actionable timepoints. While prior work has applied machine learning to EEG data to predict post-discharge outcomes after cardiac arrest,^16,23 little work has been done to identify or predict in-hospital complications that may impact outcomes, such as diffuse cerebral edema. An automated EEG-based machine learning algorithm could enable non-invasive, real-time monitoring of developing cerebral edema, a capability currently unavailable without advanced, commercially available systems. Standard care depends on intermittent and subjective reviews by epileptologists, often delayed between data acquisition, analysis, and communication to treating teams. Earlier, automated and accurate prediction of diffuse cerebral edema could facilitate the development of targeted therapeutics aimed at mitigating its effects and supporting informed clinical decision-making, as well as enhancing prognostication.

Our study aimed to 1) develop a sensitive and specific machine learning model that classified patients with and without diffuse cerebral edema based on EEG and 2) explored whether such a model could predict diffuse cerebral edema before radiographic recognition.

We included patients admitted between 2016 and 2024 who achieved ROSC after cardiac arrest and were enrolled in IRB-approved cardiac arrest registries. A full list of the clinical registries and their inclusion/exclusion criteria can be found in the Supplemental Methods. Eligible patients were required to have continuous EEG monitoring within the first 24 to 48 hours after arrest and to have at least one neuroimaging study in the first 10 days. Patients were excluded if they had fewer than 20 hours of usable EEG data, technical limitations affecting EEG analysis, or missing neuroimaging. Full exclusion criteria are provided in Fig. 1. The Boston Medical Center IRB (H-37760; H-37113; H-41069; H-40365) approved the study and waived informed consent because data were obtained retrospectively from existing registries and without deviation from standard care.

We extracted clinical data from the electronic health record, including demographics, cardiac arrest characteristics, seizure occurrence, neurological examinations, and outcomes. All neuroimaging obtained within 10 days of arrest was reviewed by a board-certified neurointensivist (CJO) to determine whether patients met diffuse cerebral edema criteria. Diffuse cerebral edema was defined as qualitative assessment of greater than 60% reduction in gray, white matter differentiation or diffusion restriction, sulcal or cisternal effacement, or evidence of herniation (Fig. 2).^6,15 An independent review by a second neurointensivist (DMG) of 13.4% of cases yielded complete agreement for diffuse cerebral edema. Patients with only focal loss of gray, white differentiation or disagreements between reviewers were classified as indeterminate. Patients were classified as having cerebral edema if criteria were met on any neuroimaging within 10 days of arrest and the onset time was assigned to the first qualifying scan. Continuous EEG recordings, acquired as part of routine clinical care, were exported from the Natus Neuroworks system into a de-identified electronic data format for analysis.²⁶

We selected 4- and 8-hour EEG segments beginning 24 hours after cardiac arrest as model inputs to test performance of both shorter and longer EEG durations at identifying severe cerebral edema. EEG recorded prior to this time was excluded from the primary analysis because early signals are heavily influenced by sedation, targeted temperature management, and other acute interventions and therefore may not reflect intrinsic cerebral physiology. To standardize all recordings, we applied a bipolar montage, using a 0.1–40 Hz bandpass filter to eliminate drift and noise, and resampled the signal at 100 Hz.^27,28 Artifacts were removed using a validated protocol targeting high-amplitude, low-variance, and rapid-change segments (Supplementary Table 1).²⁹ We extracted nine quantitative EEG features capturing different aspects of brain activity: Shannon Entropy, Signal Regularity, Spectral Power (Delta, Theta, Alpha, Beta), Alpha-Delta Ratio, Spike Frequency, and Burst Suppression Ratio.¹⁶ We augmented training data using controlled noise addition, channel shuffling, and smooth time masking to improve model robustness and generalizability.^30–32

Using these preprocessed EEG features, we performed two tasks: cerebral edema detection and prediction. The detection task assessed whether EEG patterns could be used to classify patients with and without diffuse cerebral edema. The prediction task evaluated whether the same features could identify cerebral edema, before it was identified radiographically.

We trained two machine learning model architectures to detect diffuse cerebral edema. We used a Transformer architecture its effectiveness in capturing long-range dependencies in time series data.^33,34 For comparison, we trained used a Long Short-Term Memory (LSTM) architecture, which has been applied to EEG data in prior work.^28,35

Models were initially trained on 84% of the cohort and restricted to patients with either diffuse cerebral edema or no cerebral edema, to ensure the training data reflected clinically confirmed cases of diffuse cerebral edema. The training set was augmented to increase the available data and equally represent both classes of cerebral edema. The remaining 16% were reserved as an independent test set. The test set was balanced with equal numbers of diffuse cerebral edema and non-cerebral edema cases.

Model training was conducted over 150 iterations, each using a different randomly sampled validation subset from the training set to optimize the model. We repeated this process across 30 different train-test splits, ensuring that each patient appeared in only one set per split to avoid data leakage.³⁶

We compared model performance across Transformer and LSTM architectures, with and without data augmentation, using Area Under the Curve (AUC), accuracy, sensitivity, and specificity (threshold 0.5). Performance differences were tested using the Wilcoxon signed rank test, and the best performing model was selected as the final model. Final model parameters for both architectures are shown in Supplementary Tables 3 and 4. An a priori power analysis was conducted: to achieve a sensitivity and specificity of 90 and 95%, respectively, to detect cerebral edema, a minimum of 121 samples was necessary.

EEG feature importance was assessed using Shapley Additive Explanations (SHAP), which estimates the contribution of each input feature to model predictions. SHAP values were averaged across the 30 replications to generate a stable ranking of feature influence.

Sensitivity analyses evaluated model performance in patients with in-hospital versus out-of-hospital cardiac arrest, and using EEG within the first 24 hours after arrest. In a secondary analysis, the best performing model was retrained with indeterminate cases classified as no cerebral edema to assess model performance with label uncertainty.

We demonstrated that a machine learning model can detect diffuse cerebral edema using as little as 4 hours of EEG data post-cardiac arrest. We also present preliminary evidence that an automated algorithm can predict cerebral edema onset before radiographic recognition.

Machine learning in EEG has growing applications in neuroprognostication and diagnosis, including prediction of neurologic outcomes after cardiac arrest.^16,23,28,38 Prior work, including our own, leveraged LSTM models to extract prognostic data from EEG and electrooculography (EOG),^16,35,39 but LSTMs are limited in capturing long-range dependencies in time series data. In contrast, Transformer models leverage self-attention and positional encoding to better understand patterns in time-series data.^33,34

The best-performing model in our study used a Transformer architecture with data augmentation, which created more variety in the training examples, increasing model flexibility and reducing overfitting. The Transformer's ability to focus on important parts of the data (self-attention) allowed it to make good use of this variety and better recognize key temporal patterns.

We found that 4-hour EEG segments yielded strong model performance in both detecting and predicting diffuse cerebral edema, suggesting that even brief recordings can provide meaningful information. However, 8-hour segments performed best in the prediction task, likely because they captured a broader range of EEG changes that occur as cerebral edema develops, potentially before clinical or radiographic recognition.

The most informative EEG features in both tasks were increased Delta Power, decreased Shannon Entropy, and greater Signal Regularity. Alongside prior evidence demonstrating the utility of these features in neuroprognostication,^16,22,23 our study provides support for these EEG features, which are indicative of global cortical suppression and seen in burst-suppression physiology, as robust biomarkers of cerebral edema. Subtle differences observed in relative feature importance across different EEG segment durations likely reflect evolving pathophysiology of diffuse cerebral edema and time-dependent¹⁶ susceptibility to clinical confounders, underscoring the need to consider temporal dynamics when developing real-time prognostic tools.

Subgroup analyses showed that model performance was similar, regardless of arrest location. Model performance decreased when using early EEG segments obtained within the first 24 hours after cardiac arrest, underscoring the importance of considering the timing of EEG acquisition. This reduction is likely due to confounders such as sedation, hemodynamic instability, and acute interventions. Model performance also declined when indeterminate cases were included, suggesting reduced applicability in scenarios where cerebral edema is focal, or evolving.

Continuous EEG is a widely used, guideline-recommended test in cardiac arrest.^40–42 Our findings support its potential for real-time, non-invasive detection of neurologic worsening. Early EEG-based recognition of cerebral edema prior to current standards, which may be clinical or radiographic depending on patient or institution, could enable timelier interventions compared to the current standard of care, improving patient management and prognostication.

This study has several limitations including the retrospective, single-center design with limited sample size, which may have reduced generalizability and increased the risk of overfitting. Post-hoc EEG artifact removal improved training performance but may have reduced real-world applicability where artifacts are unavoidable. Although data augmentation enhances model robustness, it may have biased features by creating artificial noise not present in native EEG data. Sedation, neuromonitoring interventions, seizures, and circadian dysrhythmia could have influenced EEG data despite starting analysis at 24 hours post-cardiac arrest. Notably, some patients experienced seizures prior to radiographic cerebral edema detection. The relationship between post-arrest seizures and subsequent cerebral edema development remains unclear and requires further research. EEGs may have had fully suppressed backgrounds, however background suppression is non-specific for cerebral edema and may be attributed to overall poor outcome,²² or sedative or antiepileptic drugs.⁴³

Neuroimaging timing varied by clinical discretion, which may have introduced cerebral edema misclassification. For patients with diffuse cerebral edema on the first post-admission scan, exact edema onset cannot be determined and may have occurred before, during, or after EEG acquisition. Only one case used in the prediction model had diffuse cerebral edema identified on the first scan, suggesting minimal impact on model performance. Limited sample size prevented matching on initial rhythm or time to ROSC, which have known associations with hypoxic-ischemic brain injury. Because clinical “edema detection” is limited either by neuroimaging intervals, or neurologic deterioration (which we did not have access to) EEG segments used in models may reflect ongoing edema that preceded imaging. However, we submit that despite that limitation, identification of edema on EEG remains an important development, as current standards to detect edema rely on late-stage clinical changes or radiographic evidence, obtained at heterogenous intervals, as in our study.

Despite these limitations, we developed the first model that we are aware of that detects and predicts cerebral edema with over 80% AUC in patients with post-cardiac arrest EEG. The implications of real-time identification can enable earlier recognition of neurologic worsening, potential future therapeutic trials, and broader applications across other acute brain injuries that experience cerebral edema. Future studies should validate these findings prospectively in multi-center cohorts with standardized EEG protocols, medication tracking, and imaging criteria, to enhance the model’s clinical utility.

We developed and validated a Transformer-based model that detects and predicts diffuse cerebral edema in post-cardiac arrest patients using EEG with high AUC, accuracy, sensitivity, and specificity. Transformer architectures consistently outperformed LSTMs, and data augmentation improved performance. These findings suggest that EEG-based machine learning may enable earlier, non-invasive identification of diffuse cerebral edema. Future studies are necessary to validate our findings.

Figure 1. Inclusion and Exclusion Criteria

Abb: EEG: electroencephalogram

a: Registries included MOCA (N = 27); MOCHA (N = 253); NEBD (N = 84): and Oculography (N = 72). Full titles and inclusion and exclusion criteria for the clinical registries listed can be found in the Supplementary Methods. Some patients were included in multiple registries.

b:Primary Cohort refers to the 124-patient cohort used for primary detection models.

c: Secondary Cohort refers to the 149-patient cohort used to retrain the primary model including indeterminate patients

Figure 2. Manual Cerebral Edema Classifications

Figure 3. Schematic of cerebral edema detection and prediction using EEG by Transformer and LSTM models

Abb: AUC: Area Under Curve; Burst Supp: Burst Suppression; EEG: Electroencephalography; Hrs: Hours; IHCA: In Hospital Cardiac Arrest; LSTM: Long Short-Term Memory; OHCA: Out of Hospital Cardiac Arrest; Shan Ent: Shannon Entropy; Spike Freq: Spike Frequency.

Figure 4. Cerebral Edema Detection Model Performance and Feature Importance

Abb: LSTM: Long Short-Term Memory; Hrs: Hours; Reg: Regularity, Shan Ent: Shannon Entropy; Spike Freq: Spike Frequency; Burst Supp: Burst Suppression

Figure 5. Cerebral Edema Prediction Model Performance and Feature Importance

Abb: Reg: Regularity, Shan Ent: Shannon Entropy; Spike Freq: Spike Frequency; Burst Supp: Burst Suppression; Hr: Hour