Automated detection of immune effector cell-associated neurotoxicity syndrome (ICANS) via quantitative EEG

Abstract

This dataset and accompanying code support the reproduction of results reported in Eckhardt, Sun, Malik, et al., "Automated detection of immune effector cell-associated neurotoxicity syndrome via quantitative EEG," Ann Clin Transl Neurol 2023;10(10):1776–1789 (PMID 37545104). In that study, we developed the EEG immune effector cell-associated neurotoxicity syndrome (E-ICANS) score, a machine learning–derived metric for automated detection of the presence and severity of neurotoxicity among patients receiving chimeric antigen receptor (CAR)-T cell therapy. The model was trained and evaluated on a retrospective cohort of CAR-T recipients monitored with continuous EEG at two tertiary care centers (2016–2020) and demonstrated high criterion, construct, and predictive validity. The dataset includes quantitative EEG features, clinical variables, neurotoxicity grades, and associated outcomes. Code for model training, evaluation, and figure generation is provided to enable full reproduction of the published analyses.

---

Background

Chimeric antigen receptor-T (CAR-T) cell therapy is a powerful treatment for relapsed and refractory hematologic malignancies, but a substantial proportion of patients—estimated at 20–70%—develop immune effector cell-associated neurotoxicity syndrome (ICANS), a clinical syndrome characterized by decreased arousal and focal neurological deficits that can progress to cerebral edema and death. Early and accurate detection of ICANS is critical for clinical management: cytokine release syndrome (CRS) often co-occurs with ICANS, yet tocilizumab, a standard CRS treatment, may worsen ICANS, while dexamethasone, the mainstay treatment for ICANS, may reduce the antitumor efficacy of CAR-T cells with prolonged use. Despite this clinical urgency, no objective, easily repeatable biomarker of ICANS has been established. Serum inflammatory markers such as CRP, ferritin, and LDH are elevated during CRS and therefore lack specificity for ICANS; cerebrospinal fluid analysis is invasive; and neuroimaging is typically normal except in the most severe cases. As a result, ICANS diagnosis relies on a standardized bedside neurological exam that is subjective, resource-intensive, and unable to capture the dynamic fluctuations in neurotoxicity severity that occur over hours.

Electroencephalography (EEG) has shown promise as a physiologic biomarker, with prior qualitative studies demonstrating that patients with ICANS exhibit increased delta and theta slowing and generalized periodic discharges correlated with worsening neurotoxicity. However, these studies were limited by small sample sizes and qualitative methods that preclude automated, scalable deployment. This study was designed to address these gaps by developing and validating the EEG ICANS (EICANS) score—an automated, quantitative, machine learning–derived metric for detecting the presence and severity of ICANS from continuous EEG data.

 

---

Methods

This was a multi-center, retrospective observational cohort study of adult inpatients who underwent continuous EEG monitoring during hospitalization for CD19- or B-cell maturation antigen (BCMA)-targeted CAR-T cell therapy at Massachusetts General Hospital (MGH) and Brigham and Women's Hospital (BWH) from May 2016 to November 2020. The study was performed under IRB approval with a waiver of written informed consent. Daily ICANS grades were determined for each patient through chart review according to ASTCT consensus criteria. Clinical variables—including laboratory values, vital signs, medications, and outcomes—were extracted from the electronic health record.

EEG recordings were obtained using Ag/AgCl scalp electrodes in the standard international 10–20 system. Raw EEG signals were resampled to 200 Hz, re-referenced to a bipolar montage, normalized to zero mean, notch filtered at 60 Hz, and bandpass filtered from 0.5 to 40 Hz. For each hospital day, a 10-minute EEG block was algorithmically selected by identifying the block with the highest mean frontal alpha band power (8–12 Hz) to preferentially sample wakefulness and avoid sleep-related features. Blocks with greater than 10% artifact were excluded, and the selected block was truncated to 9 minutes of artifact-free data. Each 9-minute block was then divided into 10-second segments with 8 seconds of overlap, and artifact segments were removed.

For each EEG sample, 94 quantitative features were extracted: 89 features in the time and frequency domains computed from four frontal leads (including spectral power in standard frequency bands, band-power ratios, coherence, spectral entropy, spectral edge frequencies, FOOOF aperiodic parameters, and time-domain features such as the Hurst exponent and Hjorth parameters), plus 5 features derived from a convolutional neural network (CNN)-based model of interictal spectrum patterns applied to 16 bipolar leads (capturing probabilities of generalized periodic discharges, generalized rhythmic delta activity, lateralized periodic discharges, lateralized rhythmic delta activity, and non-interictal-ictal continuum patterns). The top 20 features for each training fold were selected using the Maximum Relevance–Minimum Redundancy (MRMR) algorithm, with further reduction via LASSO regularization. As a control dataset, 123 age- and sex-matched EEGs with normal clinical interpretations from the same institutions were included.

The machine learning model used a Learning-to-Rank (LTR) framework, which transforms the ordinal ICANS classification problem into pairwise binary classification via logistic regression. The model was trained using nested five-fold cross-validation, with strict patient-level separation between training (80%) and test (20%) folds and stratification by ICANS grade. The model output was converted into a probability distribution over ICANS grades 0–4, and the final EICANS score was computed as the probability-weighted sum of these grades. All reported performance metrics were computed on the aggregated held-out test folds. Confidence intervals were generated via bootstrapping (1,000 iterations). Model validity was assessed through correlation of EICANS with clinical ICANS grade (criterion validity), with clinical variables known to be associated with ICANS such as ferritin, platelets, and dexamethasone use (construct validity), and with the duration of ICANS (predictive validity).

---

Data Description

The cohort consisted of 123 adult patients who received CAR-T cell therapy and underwent continuous EEG monitoring during hospitalization. An initial 136 patients were identified, with 13 excluded due to EEG being performed only before infusion (n=1), inability to locate raw EEG files or extract sufficient artifact-free data (n=9), inability to obtain a clinical exam off sedation (n=2), or encephalopathy attributed to a non-ICANS cause more than 30 days post-infusion (n=1). The median patient age was 64 years (IQR 54–69), with 79 males (64.2%) and 44 females (35.7%). The cohort was predominantly White (90.2%), with 3.3% Asian, 1.7% Black, and 4.8% other or unknown race; 92.7% were non-Hispanic. The most common underlying malignancy was diffuse large B-cell lymphoma (DLBCL; 86.2%), followed by primary mediastinal B-cell lymphoma (4.1%), follicular lymphoma (3.3%), mantle cell lymphoma (1.6%), B-cell acute lymphoblastic leukemia (1.6%), and marginal zone lymphoma (0.8%). Overall, 95.1% had aggressive histologies.

Of the 123 patients, 45 (36.6%) experienced mild to moderate ICANS (grade ≤2) and 78 (63.4%) experienced severe ICANS (grade >2) during hospitalization. The median duration of ICANS was 10 days (IQR 5–17) overall, with significantly longer duration among patients with severe ICANS (13 days, IQR 9–19) compared to mild/moderate ICANS (5 days, IQR 3–8). The median length of hospital stay was 21 days (IQR 17–30). At one year post-discharge, 43 patients (35.0%) were deceased, 57 (46.3%) were alive, and 23 (18.7%) had unknown vital status. All patients were receiving an antiseizure medication at the time of EEG monitoring, with 99.2% on levetiracetam; 3.25% received propofol and 8.94% received a benzodiazepine on the day of EEG.

From the 123 patients, 286 EEG samples were extracted (one per hospital day, with a maximum of five per patient; median one sample per patient). These were supplemented with 123 age- and sex-matched normal control EEGs, yielding a total of 409 EEG samples used in model training and evaluation. Among the non-control EEG samples, 184 (64.3%) were recorded during mild to moderate ICANS and 102 (35.7%) during severe ICANS, with a median ICANS grade of 2 (IQR 2). The mean day of EEG recording relative to CAR-T cell infusion was 11.05 (±SD 6.25), and relative to the first day of ICANS was 5.00 (±SD 5.28). Each EEG sample consists of 9 minutes of artifact-free continuous EEG data, represented as 94 quantitative features per sample. The dataset also includes daily ICANS grades, patient demographic and clinical characteristics, laboratory values, medication data, and clinical outcomes.

---

Usage Notes

Dataset Structure

The dataset is organized into the following directory structure:

s3://bdsp-opendata-credentialed/icans/
│
├── AllContinuousEEGs/
│   └── AllFullEEGs/       310 continuous EEG recordings (.mat) — 415.3 GB
│
├── EEGSegs/
│   ├── ICANSSegsInModel/  290 segments used in final model — 61.9 GB
│   ├── ICANSSegsNotPickedForFive/  25 additional segments — 5.5 GB
│   └── NormSegs/          123 age/gender-matched controls — 11.7 GB
│
└── Spreadsheets/              Metadata and feature files
    ├── ICANSFiles.xlsx
    ├── EEGFeatures.xlsx
    ├── ClinicalFeaturesEICANS.xlsx
    ├── DailyICANSData.xlsx
    └── BootstrapsForPublication/  (7 files)

 

Dataset Contents Summary

Category	Files	Size	Description
Continuous EEGs	310	415.3 GB	Full-length recordings (4–13 hours each)
ICANS Segments (In Model)	290	61.9 GB	15-min segments for training/testing (≤5 per patient)
ICANS Segments (Excluded)	25	5.5 GB	Additional segments not selected for modeling
Control Segments	123	11.7 GB	Normal EEGs from age/gender-matched controls
Metadata & Features	5	2.4 MB	Spreadsheets with patient IDs, features, clinical data
Bootstrap Results	7	9.5 MB	Model outputs from 1,200 bootstrap iterations
TOTAL	760	~494 GB	Complete E-ICANS dataset

 

EEG File Specifications

Continuous EEG Files (.mat)

• Channels: 19–21 EEG channels
• Sampling Rate: 200 Hz
• Duration: Variable (typically 4–13 hours per recording)
• Fields: data, Fs, start_time, BDSPPatientID, StudyID

EEG Segment Files (.mat)

• Channels: 4 frontal bipolar channels (Fp1-F3, Fp1-F7, Fp2-F4, Fp2-F8)
• Sampling Rate: 200 Hz
• Duration: 15 minutes per segment
• Fields: Data (N epochs × 4 channels × 2000 samples), Start, ICANS (0–4), ICE (0–10), StudyID, SourceFile

Important Notes

Patient Identification

⚠️ BDSP patient IDs embedded in EEG .mat files are INCORRECT. Always reference ICANSFiles.xlsx for correct patient mapping between StudyID and BDSPPatientID. StudyID is the primary identifier used throughout the analysis.

Data Quality

• All EEG segments have undergone artifact detection and quality filtering
• Segments were selected to have ≥60% artifact-free epochs
• One source file from ICANSSegsInModel is referenced in metadata but unavailable in the dataset

Control Group

• Control subjects (NormSegs) are assigned ICANS=0, ICE=10 by definition
• Matched to ICANS patients by age and gender
• StudyID ≥ 411 → Control subjects
• StudyID < 411 → ICANS patients

 

Recommended Download Strategy

For model development and analysis (~79 GB)

aws s3 sync s3://bdsp-opendata-credentialed/icans/EEGSegs/ ./data/EEGSegs/ --profile opendata

aws s3 sync s3://bdsp-opendata-credentialed/icans/Spreadsheets/ ./data/Spreadsheets/ --profile opendata

Includes all processed EEG segments and metadata files needed to train and evaluate models.

 

For reproducing published results (~10 MB)

aws s3 sync s3://bdsp-opendata-credentialed/icans/Spreadsheets/ ./data/Spreadsheets/ --profile opendata

Use EEGFeatures.xlsx to skip feature engineering, and bootstrap results to compare with published findings.

 

For re-processing from raw data (494 GB)

aws s3 sync s3://bdsp-opendata-credentialed/icans/ ./data/ --profile opendata

Download everything if you need to re-run the complete pipeline from continuous EEG recordings.

 

Key Spreadsheet Files

ICANSFiles.xlsx — Essential for all analyses

• Correct mapping between StudyID and BDSPPatientID
• Clinical metadata for each EEG recording
• Must be used for patient identification

EEGFeatures.xlsx — Pre-computed features

• 100 features per EEG segment (94 handcrafted + 6 CNN features)
• Can bypass the feature engineering pipeline
• Ready for direct model training

ClinicalFeaturesEICANS.xlsx — Clinical data

• Patient demographics (age, gender)
• CAR-T treatment information
• Clinical outcomes

DailyICANSData.xlsx — Longitudinal tracking

• Daily ICANS and ICE score assessments
• Temporal progression of neurotoxicity
• Treatment response tracking

BootstrapsForPublication/ — Model outputs

• Results from 1,200 bootstrap iterations
• Model coefficients, predictions, and performance metrics
• For reproducing published results

 

Code Repository

Complete analysis code and documentation available at:

GitHub: https://github.com/bdsp-core/E-ICANS

The repository includes:

• EEG preprocessing and segmentation scripts
• Feature engineering pipeline (94 handcrafted + 6 CNN features)
• Learning-to-Rank model implementation
• Visualization and analysis code
• Detailed usage instructions

---

Ethics

The study was performed under a protocol approved by the MGH Institutional Review Board (protocol # 2013P001024) using a waiver of written informed consent for analysis of data obtained as part of routine medical care.

---

Conflicts of Interest

Dr. Westover is a co-founder, scientific advisor, consultant to, and has personal equity interest in Beacon Biosignals.

---

References

1. Eckhardt CA, Sun H, Malik P, Quadri S, Santana Firme M, Jones DK, van Sleuwen M, Jain A, Fan Z, Jing J, Ge W, Danish HH, Jacobson CA, Rubin DB, Kimchi EY, Cash SS, Frigault MJ, Lee JW, Dietrich J, Westover MB. Automated detection of immune effector cell-associated neurotoxicity syndrome via quantitative EEG. Ann Clin Transl Neurol. 2023 Oct;10(10):1776-1789. doi: 10.1002/acn3.51866. Epub 2023 Aug 6. PMID: 37545104; PMCID: PMC10578889.

---