acceptodds
Under review as a conference paper at ICLR 2027

IESSEEG: An Open EEG Dataset Towards Better Understanding of Infantile Epileptic Spasms Syndrome

Abstract

We introduce IESSEEG, the first open electroencephalography (EEG) dataset dedicated to infantile epileptic spasms syndrome (IESS). IESSEEG comprises approximately 3,172 hours of EEG from 100 pediatric participants, together with diagnosis and treatment-response labels, clinician annotations, Burden of Amplitudes and Epileptiform Discharges (BASED) scores, and clinical metadata. We establish shared benchmarks for diagnosis and treatment-response prediction, comparing traditional quantitative EEG (qEEG) methods with eight EEG foundation models. Both traditional qEEG methods and foundation models achieve strong performance on the diagnosis task, while treatment-response prediction remains more challenging and benefits less consistently from fine-tuning. Further analyses suggest that combining qEEG features, learned representations, and clinical metadata can improve treatment-response prediction. By making these data available with shared benchmarks, IESSEEG enables the community to investigate EEG characteristics associated with IESS and develop more effective methods for diagnosis and treatment-response prediction.

Primary area:
datasets and benchmarks
Keywords:
electroencephalography (EEG), infantile epileptic spasms syndrome (IESS), pediatric epilepsy, datasets and benchmarks, spatiotemporal time series, treatment response prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_KC9xOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Addressing the need for EEG analysis in Infantile Epileptic Spasms Syndrome (IESS), this paper introduces the IESSEEG open dataset. The dataset comprises 150 raw long-term EEG recordings from 100 pediatric subjects—totaling approximately 3,172 hours—and includes diagnostic labels, treatment response labels, clinician annotations, BASED scores, and relevant clinical metadata. Based on this resource, the authors established benchmarks for IESS diagnosis and treatment response prediction, systematically comparing traditional quantitative EEG (qEEG) methods with eight EEG foundation models; they further analyzed the complementary roles of full-recording qEEG features and clinical metadata in predicting treatment response. The study offers significant clinical research value and contributes a valuable data resource. However, there remain certain issues that require further refinement.

Critical strengths
  1. The research motivation is clear.
  2. The experimental design is rigorous.
  3. It offers clinical research value and contributes data resources.
Critical weaknesses
  1. The abstract mentions that combining qEEG, foundation model features, and clinical metadata can improve the prediction of treatment response. It is recommended to include a key quantitative metric here to enhance the abstract's objective persuasiveness and scientific appeal.
  2. In Table 1, the median age at EEG acquisition was 0.93 years for the case group and 1.52 years for the control group, with age ranges of 0.37–8.31 years and 0.10–17.73 years, respectively; this indicates a noticeable discrepancy. Given that background rhythms, spectral distribution, and neurodevelopmental characteristics in pediatric EEGs change significantly with age, this age mismatch might lead the diagnostic model to learn age-related features rather than patterns specifically indicative of IESS abnormalities. The authors are advised to conduct age-matching or age-stratified analyses and compare diagnostic performance before and after matching to assess potential age-related confounding effects in the dataset.
  3. Section 3.3 describes how seven clinicians independently scored 20 raw long-term EEG recordings, resulting in a total of 140 BASED scoring records. However, the proportion of records with BASED scores is relatively low compared to the 150 raw long-term EEG recordings in the dataset. The authors should clarify the criteria used to select these 20 recordings, the distribution of BASED score grades, and the changes in scores before and after treatment. Additionally, reporting inter-rater reliability would better demonstrate the quality of this annotated resource and clarify whether it is best suited for model validation, studies on scoring consistency, or training automated BASED scoring models.
  4. The study employed a fixed 5-fold cross-validation approach. Given the limited dataset size—comprising only 100 patients, with the response group containing just 50 individuals (28 vs. 22)—it is recommended to perform multiple splits using different random seeds or to employ bootstrapping for confidence interval estimation to ensure more robust statistical results.
  5. The conclusion summarizes the study's contributions regarding the construction of the IESSEEG open dataset and the prediction of treatment response; however, the discussion of study limitations lacks sufficient focus. It is recommended that the author add a dedicated discussion on limitations at the end of Chapter 6, or supplement the current limitations of IESSEEG in Chapter 7, to provide direction for future data collection aimed at generalization within the field.
Questions for authors

None

6 comments
AuthorsOpenReview ↗

Response to Reviewer KC9x (1/4)

We thank the reviewer for the detailed assessment and constructive suggestions. In response, we repeated the diagnostic benchmark in the age-matched subgroup and added patient-level bootstrap analysis. We also report the BASED score distribution and inter-rater reliability, and provide content we will add to the abstract and limitations discussion.

Critical Weaknesses 1 – Quantitative result in the abstract

We will add the following content to the abstract:

Further analyses suggest that qEEG features and clinical metadata can improve treatment-response prediction from learned representations. For LaBraM, adding whole-recording qEEG increased pre-treatment response prediction AUROC from 67.6% to 72.1%. Adding the interval from spasm onset to the first EEG further increased AUROC to 74.3%.

Critical Weaknesses 2 – Diagnosis in the age-matched subgroup

Diagnostic discrimination remained high in the age-matched subgroup. Fully fine-tuned foundation models achieved mean five-fold AUROCs of 91.7–98.1% against patient diagnosis labels and 89.9–96.0% against clinician diagnosis labels. We evaluated the age-matched subgroup of 40 cases and 20 controls in the metadata, using the same benchmark settings as in Section 4 and Appendix B.

For the age-matched subgroup, median acquisition age is 0.91 years (interquartile range 0.64–1.81) for cases and 0.83 years (0.62–1.95) for controls. Mean ages are 1.72 years in both groups. Table 1 compares the full-cohort and subgroup results against both patient and clinician diagnosis labels.

Table 1. IESS diagnosis task results against patient diagnosis labels. AUROC, balanced accuracy and 95% confidence intervals are calculated as in Section 4.4 and Appendix B.6. The full cohort has 100 patients (50 cases and 50 controls) and 200 routine clips. The age-matched subgroup has 60 patients (40 cases and 20 controls) and 120 routine clips.

Patient diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet91.4 [87.3–95.5]85.5 [80.1–90.9]94.2 [90.5–98.0]80.0 [72.3–87.7]
Expanded qEEG93.1 [88.8–97.4]85.5 [79.7–91.3]93.1 [87.9–98.4]83.1 [75.9–90.4]
Frozen encoders
BIOT87.6 [81.5–93.7]81.5 [75.5–87.5]85.6 [76.7–94.6]76.9 [67.5–86.3]
LaBraM94.8 [91.3–98.3]86.5 [80.6–92.4]94.2 [89.2–99.3]81.3 [73.9–88.6]
CBraMod90.7 [85.3–96.1]84.5 [78.7–90.3]92.5 [86.5–98.5]81.9 [74.9–88.9]
EEGPT94.7 [90.9–98.4]87.5 [81.7–93.3]94.5 [89.4–99.7]84.4 [76.4–92.4]
LUNA87.5 [81.0–93.9]84.0 [78.1–89.9]88.3 [80.8–95.7]82.5 [75.3–89.7]
REVE92.8 [88.2–97.3]87.5 [82.1–92.9]95.6 [91.0–100.0]88.8 [82.8–94.7]
CodeBrain94.6 [91.0–98.1]88.0 [82.2–93.8]91.6 [85.5–97.7]81.3 [74.1–88.4]
CSBrain87.9 [82.1–93.7]80.5 [74.4–86.6]85.2 [76.9–93.4]78.1 [70.0–86.3]
Fully fine-tuned encoders
BIOT97.2 [94.9–99.4]90.5 [86.0–95.0]91.7 [85.5–97.9]86.9 [77.9–95.9]
LaBraM97.6 [95.0–100.0]93.5 [89.2–97.8]96.6 [93.1–100.0]88.8 [80.7–96.8]
CBraMod96.3 [92.7–99.9]91.0 [86.1–95.9]97.8 [95.6–100.0]91.3 [85.7–96.8]
EEGPT95.4 [91.8–99.0]91.0 [86.3–95.7]94.2 [88.9–99.5]79.4 [69.1–89.6]
LUNA93.9 [89.0–98.8]87.0 [81.5–92.5]91.7 [85.9–97.6]80.0 [69.5–90.5]
REVE94.8 [90.8–98.7]88.5 [83.0–94.0]94.8 [89.9–99.7]88.1 [79.9–96.3]
CodeBrain96.1 [92.9–99.3]90.5 [85.6–95.4]98.1 [96.1–100.0]85.0 [75.9–94.1]
CSBrain96.5 [93.6–99.4]91.0 [86.4–95.6]95.8 [91.4–100.0]90.6 [83.9–97.4]
AuthorsOpenReview ↗

Response to Reviewer KC9x (2/4)

Critical Weaknesses 2 – Diagnosis in the age-matched subgroup (continued)

Table 1. IESS diagnosis task results against clinician diagnosis labels. AUROC, balanced accuracy and 95% confidence intervals are calculated as in Section 4.4 and Appendix B.6. The full cohort has 100 patients (50 cases and 50 controls) and 200 routine clips. The age-matched subgroup has 60 patients (40 cases and 20 controls) and 120 routine clips.

Clinician diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet90.7 [85.9–95.4]83.5 [77.9–89.2]93.9 [88.7–99.2]78.1 [70.4–85.8]
Expanded qEEG94.7 [91.3–98.1]86.2 [80.7–91.7]97.0 [93.7–100.0]86.7 [80.3–93.1]
Frozen encoders
BIOT88.4 [82.9–94.0]81.5 [75.7–87.3]88.4 [81.3–95.6]77.6 [69.2–85.9]
LaBraM97.3 [95.3–99.4]89.2 [83.9–94.4]95.5 [91.7–99.3]85.2 [78.7–91.6]
CBraMod92.2 [87.8–96.6]85.1 [79.6–90.6]92.6 [86.9–98.3]81.7 [74.3–89.2]
EEGPT93.2 [88.9–97.5]84.7 [78.5–90.8]91.5 [84.5–98.5]81.9 [72.7–91.1]
LUNA88.1 [81.6–94.6]84.7 [78.8–90.5]89.2 [82.4–96.0]84.5 [76.8–92.1]
REVE91.1 [86.3–96.0]85.5 [79.9–91.1]94.2 [88.8–99.7]85.8 [77.8–93.8]
CodeBrain95.1 [91.9–98.4]89.9 [84.7–95.1]93.2 [87.8–98.6]83.1 [75.3–90.9]
CSBrain90.2 [84.8–95.6]81.1 [75.2–86.9]89.2 [82.0–96.4]81.7 [74.1–89.4]
Fully fine-tuned encoders
BIOT95.0 [91.7–98.2]86.6 [81.2–92.0]91.7 [86.2–97.2]81.8 [72.5–91.1]
LaBraM97.4 [94.9–99.9]92.5 [88.2–96.7]92.9 [86.6–99.2]85.3 [76.3–94.2]
CBraMod95.2 [91.4–99.0]90.0 [85.4–94.6]94.7 [89.4–100.0]89.1 [82.3–96.0]
EEGPT92.6 [88.1–97.1]87.0 [81.4–92.6]91.1 [83.6–98.6]80.3 [70.9–89.8]
LUNA92.4 [87.2–97.7]85.8 [80.0–91.6]89.9 [83.1–96.7]78.3 [68.3–88.3]
REVE91.4 [86.6–96.3]84.7 [78.6–90.8]93.4 [88.1–98.8]83.1 [74.3–92.0]
CodeBrain96.2 [93.0–99.5]90.5 [85.9–95.2]96.0 [92.1–99.9]83.8 [75.0–92.7]
CSBrain96.4 [93.2–99.6]92.7 [88.8–96.7]94.9 [89.5–100.0]88.7 [81.2–96.3]

Critical Weaknesses 3 – BASED coverage, score distribution and reliability

The BASED annotations in IESSEEG come from the reliability study by Swarz et al. [1] (Section 3.3). In the study, ten pre-treatment and ten post-treatment long EEG recordings from 20 different patients were randomly selected for annotation. Four experts and three trainees independently reviewed all 20 recordings and assigned a BASED score to each, yielding 140 ratings. Scoring followed the 2021 BASED criteria [2, 3].

Please refer to Table 2 for the distribution of released BASED scores. Pre-treatment ratings are concentrated at grade 5, whereas post-treatment ratings are more dispersed. Because these 20 recordings come from different patients, the table compares pre-treatment and post-treatment score distributions across groups. Within-patient changes after treatment cannot be assessed.

Table 2. BASED grade counts across seven ratings per recording.

Recording groupBASED score 0BASED score 1BASED score 2BASED score 3BASED score 4BASED score 5Total
Pre-treatment003485570
Post-treatment28811106770

Following Mytinger et al. [3], we computed the intraclass correlation coefficient (ICC) from all 140 original ratings. The ICC was 0.78 (95% confidence interval 0.65–0.89) across all seven raters, 0.85 (0.73–0.93) for experts and 0.68 (0.45–0.84) for trainees. Both the overall ICC of 0.78 and the expert-rater ICC of 0.85 fall within the good-agreement range. These values are comparable to the ICC of 0.86 reported in the recent multicenter BASED validation study by Mytinger et al. [3], supporting the reliability of our BASED annotations. We provide these annotations as a supplementary resource to support exploratory research and analysis by IESSEEG users and the broader research community.

AuthorsOpenReview ↗

Response to Reviewer KC9x (3/4)

Critical Weaknesses 4 – Confidence intervals and bootstrap analysis

We added patient-level bootstrap results and confidence intervals for diagnosis and response prediction. Keeping the experimental settings unchanged, we resampled patients within each test fold 1,000 times. Tables 3 and 4 report bootstrap mean AUROC and balanced accuracy with 95% percentile confidence intervals.

Table 3. Diagnosis AUROC and balanced accuracy (%). Entries are bootstrap means with 95% percentile confidence intervals.

Patient diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet91.4 [87.1–95.7]85.5 [80.0–91.0]94.3 [89.7–97.8]80.3 [72.4–87.6]
Expanded qEEG93.2 [87.9–97.2]85.6 [79.7–90.8]93.3 [87.5–98.2]83.2 [76.2–89.5]
Frozen encoders
BIOT87.6 [80.5–93.6]81.5 [75.6–87.0]85.5 [74.3–94.8]76.8 [66.7–86.1]
LaBraM94.8 [91.1–98.1]86.6 [80.7–92.0]94.2 [87.9–98.8]81.2 [73.5–87.8]
CBraMod90.7 [84.7–96.0]84.5 [79.2–89.6]92.5 [85.1–98.0]81.9 [74.4–88.5]
EEGPT94.6 [89.7–98.0]87.4 [81.5–92.3]94.5 [88.3–99.0]84.1 [75.9–91.8]
LUNA87.5 [79.9–93.8]84.2 [78.2–89.4]88.2 [80.3–95.4]82.3 [75.1–89.4]
REVE92.7 [87.1–97.0]87.5 [81.9–92.4]95.5 [89.5–99.6]88.7 [82.7–94.3]
CodeBrain94.6 [90.4–97.8]88.2 [81.9–93.4]91.6 [84.4–97.6]81.2 [75.2–87.5]
CSBrain88.0 [81.3–93.4]80.6 [74.3–85.9]85.2 [76.1–93.2]78.3 [70.1–86.1]
Fully fine-tuned encoders
BIOT97.1 [94.2–99.1]90.5 [85.4–94.5]91.8 [84.0–98.0]87.0 [77.5–95.1]
LaBraM97.5 [94.6–99.6]93.4 [89.0–97.2]96.5 [91.3–99.7]88.6 [79.9–96.3]
CBraMod96.2 [92.1–99.4]90.9 [85.6–95.4]97.8 [94.7–100.0]91.2 [85.4–96.2]
EEGPT95.3 [91.0–98.5]91.0 [86.4–95.3]94.1 [87.3–99.0]79.1 [69.2–88.4]
LUNA93.8 [88.0–98.5]87.0 [81.1–92.2]91.6 [84.8–97.3]79.7 [68.7–89.5]
REVE94.6 [90.1–98.0]88.3 [82.8–93.2]94.6 [88.7–99.0]87.8 [79.3–95.4]
CodeBrain96.0 [92.5–98.7]90.4 [85.1–94.6]98.1 [95.4–100.0]84.7 [74.4–93.3]
CSBrain96.5 [93.1–99.1]91.1 [86.3–95.2]95.8 [90.4–99.9]90.4 [83.1–96.8]

Table 3. Diagnosis AUROC and balanced accuracy (%). Entries are bootstrap means with 95% percentile confidence intervals.

Clinician diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet90.6 [85.5–95.2]83.6 [78.1–89.2]94.0 [87.8–98.6]78.4 [70.7–85.9]
Expanded qEEG94.7 [90.6–97.8]86.2 [81.0–91.2]97.0 [92.6–99.8]86.9 [80.4–92.5]
Frozen encoders
BIOT88.4 [82.0–94.0]81.5 [75.6–87.1]88.2 [79.8–95.2]77.6 [68.7–85.6]
LaBraM97.4 [95.0–99.2]89.2 [83.9–93.9]95.6 [91.2–98.7]85.2 [78.2–91.1]
CBraMod92.3 [86.9–96.4]85.1 [79.7–89.9]92.4 [85.3–97.6]81.9 [74.3–88.5]
EEGPT93.0 [87.9–97.0]84.6 [78.5–90.1]91.6 [83.9–98.0]81.8 [72.9–90.0]
LUNA88.1 [80.4–94.5]84.7 [78.7–89.8]89.2 [81.7–96.2]84.5 [76.8–91.3]
REVE91.1 [85.8–95.6]85.4 [80.0–90.5]94.3 [88.2–99.3]86.0 [77.6–92.7]
CodeBrain95.1 [91.4–98.2]90.0 [84.5–95.0]93.4 [87.0–98.2]83.3 [75.2–90.3]
CSBrain90.2 [83.6–95.3]81.1 [75.3–86.6]89.2 [80.8–95.9]82.0 [74.0–88.9]
Fully fine-tuned encoders
BIOT94.9 [91.0–97.8]86.6 [81.1–91.7]91.6 [84.3–96.9]81.9 [72.5–90.8]
LaBraM97.3 [94.3–99.5]92.4 [88.3–96.0]93.0 [86.0–98.6]85.2 [76.4–93.2]
CBraMod95.0 [90.4–98.4]89.8 [85.1–94.0]94.8 [88.3–99.2]89.2 [82.4–94.9]
EEGPT92.4 [87.4–96.5]87.0 [81.9–92.0]91.1 [83.1–97.9]80.2 [71.8–89.0]
LUNA92.3 [86.1–97.0]85.8 [79.6–91.0]89.9 [82.0–96.1]78.3 [68.1–87.2]
REVE91.2 [85.7–95.5]84.5 [78.8–90.4]93.4 [87.1–98.4]83.1 [73.9–91.3]
CodeBrain96.1 [92.1–98.9]90.5 [85.2–94.9]95.9 [90.6–99.4]83.7 [74.7–91.3]
CSBrain96.2 [92.3–98.8]92.8 [88.7–96.4]94.8 [87.8–99.2]88.8 [80.6–95.5]
Reviewer_wGKvOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper releases IESSEEG, a dataset containing approximately 3172 hours of scalp EEG from 100 subjects, and establishes a benchmark for IESS diagnosis and treatment response prediction. They compare conventional qEEG with 8 EEG foundation models and explore the complementary value of whole-recording qEEG and clinical information.

Critical strengths

S1. The public release of paired pre- and post-treatment EEG, treatment outcome labels, and clinical annotations for IESS addresses the scarcity of public data in this setting and offers clear value for clinical research and reuse.

S2. Experiments cover 8 EEG foundation models under frozen and fully fine-tuned settings, alongside conventional qEEG baselines. Fixed patient-level splits and public code provide a reasonably comprehensive and reproducible basis for future comparisons.

Critical weaknesses

W1. The paper could provide more detail on data quality and clinical annotation procedures, and further assess how potential age confounding affects the results.

Questions for authors

Q1. Could the authors provide more details such as the artifact rejection rate in the current preprocessing pipeline and the duration of usable signal?

Q2. Were the diagnostic labels for the 200 routine EEG recordings assigned by a single clinician or through consensus among multiple clinicians? How were any disagreements handled?

Q3. Given that §3.1 and Table 1 indicate that cases and controls were not fully age-matched, could the authors provide an age-only baseline or diagnostic results on an age-matched subset to assess the impact of age confounding?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wGKv (1/2)

We thank the reviewer for recognizing the value of the IESSEEG dataset in providing common ground for research on IESS diagnosis and treatment response. Below, we offer further details on artifact rejection, usable signal duration and clinical annotation, together with diagnostic results from the age-matched subgroup.

Question 1, Critical Weaknesses 1 – Artifact rejection and usable signal duration

For qEEG preprocessing, we follow Smith et al. [1] and Rajaraman et al. [2] (Appendices B.1–B.2). It automatically rejects artifactual windows on a one-second basis. For the foundation models, we adopt preprocessing procedures from their official implementations. For most models, these include bandpass filtering at approximately 0.3–75 Hz, notch filtering to remove line noise, re-referencing and normalization. These downstream pipelines do not discard windows based on artifacts. We provide the model-specific preprocessing details in Appendix B.3.

Table 1. Signal retained after artifact rejection in the qEEG pipeline.

InputInput (h)Retained (h)Retention (%)
Clinician-selected EEG166.90153.6092.0
Simulated routine EEG100.0688.6588.6
Pre-treatment long EEG1134.98993.4087.5
Post-treatment long EEG1021.30868.7985.1

Question 2, Critical Weaknesses 1 – Diagnostic labels annotation process

The clinician diagnosis labels for the 200 simulated routine EEG segments were assigned through consensus among three expert clinicians (Section 3.3). All annotators were expert clinicians from the same institution with a shared clinical training background. They met before annotation to establish consensus assessment criteria. During manual visual review, they judged a segment as consistent with IESS when they identified abnormalities commonly seen in children with infantile spasms, such as epileptiform spikes, paroxysmal fast activity, high voltage or paroxysmal voltage attenuations. After annotation, they met again to discuss disagreements and agree on the final labels.

Question 3, Critical Weaknesses 1 – Diagnosis in the age-matched subgroup

Diagnostic discrimination remained high in the age-matched subgroup. Fully fine-tuned foundation models achieved mean five-fold AUROCs of 91.7–98.1% against patient diagnosis labels and 89.9–96.0% against clinician diagnosis labels. We evaluated the age-matched subgroup of 40 cases and 20 controls in the metadata, using the same benchmark settings as in Section 4 and Appendix B.

For the age-matched subgroup, median acquisition age is 0.91 years (interquartile range 0.64–1.81) for cases and 0.83 years (0.62–1.95) for controls. Mean ages are 1.72 years in both groups. Table 2 compares the full-cohort and subgroup results against both patient and clinician diagnosis labels.

AuthorsOpenReview ↗

Response to Reviewer wGKv (2/2)

Question 3, Critical Weaknesses 1 – Diagnosis in the age-matched subgroup (continued)

Table 2. IESS diagnosis task results against patient diagnosis labels. AUROC, balanced accuracy and 95% confidence intervals are calculated as in Section 4.4 and Appendix B.6. The full cohort has 100 patients (50 cases and 50 controls) and 200 routine clips. The age-matched subgroup has 60 patients (40 cases and 20 controls) and 120 routine clips.

Patient diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet91.4 [87.3–95.5]85.5 [80.1–90.9]94.2 [90.5–98.0]80.0 [72.3–87.7]
Expanded qEEG93.1 [88.8–97.4]85.5 [79.7–91.3]93.1 [87.9–98.4]83.1 [75.9–90.4]
Frozen encoders
BIOT87.6 [81.5–93.7]81.5 [75.5–87.5]85.6 [76.7–94.6]76.9 [67.5–86.3]
LaBraM94.8 [91.3–98.3]86.5 [80.6–92.4]94.2 [89.2–99.3]81.3 [73.9–88.6]
CBraMod90.7 [85.3–96.1]84.5 [78.7–90.3]92.5 [86.5–98.5]81.9 [74.9–88.9]
EEGPT94.7 [90.9–98.4]87.5 [81.7–93.3]94.5 [89.4–99.7]84.4 [76.4–92.4]
LUNA87.5 [81.0–93.9]84.0 [78.1–89.9]88.3 [80.8–95.7]82.5 [75.3–89.7]
REVE92.8 [88.2–97.3]87.5 [82.1–92.9]95.6 [91.0–100.0]88.8 [82.8–94.7]
CodeBrain94.6 [91.0–98.1]88.0 [82.2–93.8]91.6 [85.5–97.7]81.3 [74.1–88.4]
CSBrain87.9 [82.1–93.7]80.5 [74.4–86.6]85.2 [76.9–93.4]78.1 [70.0–86.3]
Fully fine-tuned encoders
BIOT97.2 [94.9–99.4]90.5 [86.0–95.0]91.7 [85.5–97.9]86.9 [77.9–95.9]
LaBraM97.6 [95.0–100.0]93.5 [89.2–97.8]96.6 [93.1–100.0]88.8 [80.7–96.8]
CBraMod96.3 [92.7–99.9]91.0 [86.1–95.9]97.8 [95.6–100.0]91.3 [85.7–96.8]
EEGPT95.4 [91.8–99.0]91.0 [86.3–95.7]94.2 [88.9–99.5]79.4 [69.1–89.6]
LUNA93.9 [89.0–98.8]87.0 [81.5–92.5]91.7 [85.9–97.6]80.0 [69.5–90.5]
REVE94.8 [90.8–98.7]88.5 [83.0–94.0]94.8 [89.9–99.7]88.1 [79.9–96.3]
CodeBrain96.1 [92.9–99.3]90.5 [85.6–95.4]98.1 [96.1–100.0]85.0 [75.9–94.1]
CSBrain96.5 [93.6–99.4]91.0 [86.4–95.6]95.8 [91.4–100.0]90.6 [83.9–97.4]

Table 2. IESS diagnosis task results against clinician diagnosis labels. AUROC, balanced accuracy and 95% confidence intervals are calculated as in Section 4.4 and Appendix B.6. The full cohort has 100 patients (50 cases and 50 controls) and 200 routine clips. The age-matched subgroup has 60 patients (40 cases and 20 controls) and 120 routine clips.

Clinician diagnosis labels

MethodFull cohort AUROCFull cohort Balanced accuracyAge-matched AUROCAge-matched Balanced accuracy
qEEG triplet90.7 [85.9–95.4]83.5 [77.9–89.2]93.9 [88.7–99.2]78.1 [70.4–85.8]
Expanded qEEG94.7 [91.3–98.1]86.2 [80.7–91.7]97.0 [93.7–100.0]86.7 [80.3–93.1]
Frozen encoders
BIOT88.4 [82.9–94.0]81.5 [75.7–87.3]88.4 [81.3–95.6]77.6 [69.2–85.9]
LaBraM97.3 [95.3–99.4]89.2 [83.9–94.4]95.5 [91.7–99.3]85.2 [78.7–91.6]
CBraMod92.2 [87.8–96.6]85.1 [79.6–90.6]92.6 [86.9–98.3]81.7 [74.3–89.2]
EEGPT93.2 [88.9–97.5]84.7 [78.5–90.8]91.5 [84.5–98.5]81.9 [72.7–91.1]
LUNA88.1 [81.6–94.6]84.7 [78.8–90.5]89.2 [82.4–96.0]84.5 [76.8–92.1]
REVE91.1 [86.3–96.0]85.5 [79.9–91.1]94.2 [88.8–99.7]85.8 [77.8–93.8]
CodeBrain95.1 [91.9–98.4]89.9 [84.7–95.1]93.2 [87.8–98.6]83.1 [75.3–90.9]
CSBrain90.2 [84.8–95.6]81.1 [75.2–86.9]89.2 [82.0–96.4]81.7 [74.1–89.4]
Fully fine-tuned encoders
BIOT95.0 [91.7–98.2]86.6 [81.2–92.0]91.7 [86.2–97.2]81.8 [72.5–91.1]
LaBraM97.4 [94.9–99.9]92.5 [88.2–96.7]92.9 [86.6–99.2]85.3 [76.3–94.2]
CBraMod95.2 [91.4–99.0]90.0 [85.4–94.6]94.7 [89.4–100.0]89.1 [82.3–96.0]
EEGPT92.6 [88.1–97.1]87.0 [81.4–92.6]91.1 [83.6–98.6]80.3 [70.9–89.8]
LUNA92.4 [87.2–97.7]85.8 [80.0–91.6]89.9 [83.1–96.7]78.3 [68.3–88.3]
REVE91.4 [86.6–96.3]84.7 [78.6–90.8]93.4 [88.1–98.8]83.1 [74.3–92.0]
CodeBrain96.2 [93.0–99.5]90.5 [85.9–95.2]96.0 [92.1–99.9]83.8 [75.0–92.7]
CSBrain96.4 [93.2–99.6]92.7 [88.8–96.7]94.9 [89.5–100.0]88.7 [81.2–96.3]

References

[1] Smith et al. (2021). Computational characteristics of interictal EEG as objective markers of epileptic spasms. Epilepsy Research, 176: 106704.

[2] Rajaraman et al. (2024). Computational EEG attributes predict response to therapy for epileptic spasms. Clinical Neurophysiology, 163: 39–46.

Thank you for your response. The rebuttal has addressed my concerns, and I will maintain my positive rating.

Thank you for your follow-up and positive feedback. We sincerely appreciate the time you have taken to review our work and discuss your concerns with us.

Reviewer_uCsmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Paper introduces a new dataset of EEG from 100 pediatric participants focused on epileptic episodes. The paper comapres foundation models and qEEG based method on classification as well as treatment response to epilespy. On the latter, there is much scope for research.

Critical strengths
  1. In my knowledge there are only a limited number of public EEG datasets, and even smaller is pediatric one. So, it is a valuable resource.
  2. The method comparisons presented shows that treatment response has much to study on.
Critical weaknesses
  1. Not much, except that it is rather small in terms number of subjects.
Questions for authors
  • Can we get some additional metadata beyond the time series data? I believe that would provide richer information set for studying the treatment response.
  • Is it possible to discuss a bit more detail on what makes treatment response a harder problem by means of, say, ablation study, or qEEG study?
5 comments
AuthorsOpenReview ↗

Response to Reviewer uCsm (1/3)

We thank the reviewer for recognizing the value of IESSEEG and for the constructive suggestions. Below, we describe the released metadata and discuss why treatment response is harder to predict.

Critical Weaknesses 1 – Dataset size

We agree that a cohort of 100 participants is limited in size. We plan to follow this release with expanded datasets and more in-depth studies of IESS.

Question 1 – Metadata and annotations

Yes. The current release (v1.1) includes additional metadata alongside the EEG (Section 3.4). We removed or obfuscated personally identifiable information before release. Patient-level metadata are provided at https://huggingface.co/datasets/Capur/IESSEEG/blob/main/participants.tsv, and recording-level metadata at https://huggingface.co/datasets/Capur/IESSEEG/blob/main/recordings.tsv. Table 1 summarizes these fields from our Hugging Face release.

Table 1. Metadata fields in IESSEEG.

Field nameMeaning
age_at_eeg_yearsAge at the pre-treatment EEG for cases or original EEG for controls (years).
genderGender.
etiologyUnderlying cause: structural abnormality, genetic abnormality, or unknown.
age_at_onset_monthsAge at spasm onset for cases (months).
lead_time_days, lead_time_monthsInterval from spasm onset to the first EEG, in days or months. lead_time_ukiss is its approximate logarithm.
treatment_categoryHormonal therapy [adrenocorticotropic hormone (ACTH) or prednisolone] without vigabatrin; vigabatrin without hormonal therapy; or combination therapy (simultaneous hormonal therapy and vigabatrin).
eeg_interval_daysDays between the pre-treatment and post-treatment long EEG recordings.
initial_response_label, treatment_response_labelInitial response at approximately two weeks and sustained response at approximately six weeks, respectively (1 = responder, 0 = non-responder).
control_suspected_spasms, control_suspected_other_seizuresControls referred for suspected infantile spasms or other seizure types, respectively.
age_matched_case, age_matched_controlIndicators for the 40 cases and 20 controls in the pre-existing age-matched subset, respectively.
sleep_awake_labelAwake or sleep state for clinician-selected segments only.
based_scoreIndividual clinician-assigned 2021 BASED score of EEG abnormality (0–5); 140 ratings from seven clinicians for 20 long recordings [1].

Individual participants' ages at EEG recording, spasm-onset ages, lead times and response labels are also listed in Appendix A (Table A1).

We agree that clinical metadata can provide useful information for studying treatment response. We also found that lead time improved response prediction (Section 6.3). Adding lead time to the combination of model representations and whole-recording qEEG features increased pre-treatment sampled-segment AUROC from 72.1% to 74.3% for LaBraM and from 66.6% to 70.9% for CBraMod.

AuthorsOpenReview ↗

Response to Reviewer uCsm (2/3)

Question 2 – Insights into response prediction

We speculate that response prediction depends on several kinds of information that current methods have difficulty capturing together. The relevant patterns might need to be recognized jointly across EEG channels and from short to very long temporal scales. Metadata are also needed as proxies for information that is difficult to infer from EEG.

Predicting treatment response is a recognized challenge in clinical research and practice. Several EEG features are associated with IESS diagnosis, but their value for predicting treatment response is less clear. For example, hypsarrhythmia, a feature associated with IESS diagnosis, was not associated with first-line response in a multicenter cohort [2]. Clinical attributes such as etiology and treatment delay have been associated with response [3], yet accurate pre-treatment prediction for individual patients remains difficult [4]. Previous studies have attributed the difficulty of studying treatment response in IESS to small cohorts and variation in underlying causes, treatments and the timing of follow-up EEG [2, 4, 5].

The longstanding lack of open IESS datasets has created a need for computational approaches that draw on knowledge beyond IESS recordings alone. Two broad computational approaches address this need. Traditional qEEG uses predefined measurements informed by physiology and existing EEG studies [4, 6]. Prior studies have linked response to functional connectivity, long-range temporal correlations and clinical metadata such as lead time [4, 7]. More recently, EEG foundation models have offered a second route through self-supervised pretraining on much larger EEG collections [8, 9, 10]. This pretraining does not require treatment-response labels or a population matched to IESS; the aim is to transfer representations learned in other cohorts to the smaller target cohort.

A central limitation of traditional qEEG, however, is its reliance on features designed in advance. Such features may miss patterns that are difficult to characterize through fixed mathematical summaries. Foundation models, on the other hand, learn representations directly from EEG waveforms [9, 10], but majority use short input contexts that limit direct modeling of temporal relationships across a long recording [12]. The two approaches draw on different perspectives and can characterize EEG at different temporal scales, so they may capture different information relevant to treatment response.

We therefore tested these two approaches in combination. Using the procedure in Section 6.2, adding whole-recording qEEG to frozen representations increases mean AUROC from 67.6% to 72.1% for LaBraM and from 57.7% to 66.6% for CBraMod, compared with 62.3% for qEEG alone. This result suggests that qEEG provides predictive information that complements the learned representations.

We next examined whether the gains from adding qEEG come from information captured over short or long EEG intervals. Keeping the other settings unchanged, we varied only the duration of EEG used to compute qEEG (Table 2). The results show an overall trend toward larger gains at longer qEEG measurement durations.

Table 2. Mean five-fold pre-treatment response AUROC (%) for foundation-model representations combined with the response qEEG. The first row gives representation-only references. Settings follow Section 6.2 and Appendix C.2, varying only qEEG measurement duration. qEEG intervals started at the sampled segment's start.

qEEG measurement durationLaBraMCBraMod
Representation alone67.657.7
5 minutes69.161.0
10 minutes69.459.6
15 minutes69.759.4
30 minutes69.759.4
1 hour70.760.7
2 hours71.061.3
5 hours71.063.3
Up to 10 hours70.765.7
Whole recording72.166.6
AuthorsOpenReview ↗

Response to Reviewer uCsm (3/3)

Question 2 – Insights into response prediction (continued)

These gains are encouraging. However, the EEG durations associated with the larger gains extend to several hours, beyond what current EEG foundation models can process as a single input. On the other hand, the information captured by longer-duration qEEG remains constrained by predefined mathematical construction. Models that can discover patterns directly from long EEG recordings could offer a more flexible way to capture response-related information.

Metadata also provide another source of information that may be difficult to infer from EEG alone. Adding lead time increases mean AUROC from 72.1% to 74.3% for LaBraM and from 66.6% to 70.9% for CBraMod (Section 6.3).

The gains from combining learned representations, longer-duration qEEG and metadata suggest that current computational methods capture useful but incomplete information about response. A model that analyzes EEG at multiple temporal scales could discover complex patterns within short windows and potentially weak and sparse signals over long recordings. It could also accept metadata as input and use auxiliary supervision informed by existing qEEG studies. Developing a model with these capabilities remains a challenge for IESS response prediction.

References

[1] Swarz et al. (2023). Reliability of the 2021 BASED score and hypsarrhythmia for assessment of infantile spasms. American Epilepsy Society Annual Meeting, abstract 2.138. https://aesnet.org/abstractslisting/reliability-of-the-2021-based-score-and-hypsarrhythmia-for-assessment-of-infantile-spasms.

[2] Demarest et al. (2017). The impact of hypsarrhythmia on infantile spasms treatment response: Observational cohort study from the National Infantile Spasms Consortium. Epilepsia, 58 (12): 2098–2103. https://doi.org/10.1111/epi.13937.

[3] Surana et al. (2020). Infantile spasms: Etiology, lead time and treatment response in a resource limited setting. Epilepsy & Behavior Reports, 14: 100397. https://doi.org/10.1016/j.ebr.2020.100397.

[4] Rajaraman et al. (2024). Computational EEG attributes predict response to therapy for epileptic spasms. Clinical Neurophysiology, 163: 39–46.

[5] Mostaghimi et al. (2026). Prediction of treatment response in infantile epileptic spasms syndrome using EEG phase-amplitude coupling. Epilepsia Open, 11 (3): 801–813. https://doi.org/10.1002/epi4.70230.

[6] Smith et al. (2021). Computational characteristics of interictal EEG as objective markers of epileptic spasms. Epilepsy Research, 176: 106704.

[7] Shrey et al. (2018). Strength and stability of EEG functional connectivity predict treatment response in infants with epileptic spasms. Clinical Neurophysiology, 129 (10): 2137–2148.

[8] Yang et al. (2023). BIOT: Biosignal transformer for cross-data learning in the wild. In Advances in Neural Information Processing Systems, volume 36, pp. 78240–78260. https://doi.org/10.52202/075280-3420. https://proceedings.neurips.cc/paper_files/paper/2023/hash/f6b30f3e2dd9cb53bbf2024402d02295-Abstract.html.

[9] Jiang et al. (2024). Large brain model for learning generic representations with tremendous EEG data in BCI. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2024/hash/47393e8594c82ce8fd83adc672cf9872-Abstract-Conference.html.

[10] Wang et al. (2025). CBraMod: A criss-cross brain foundation model for EEG decoding. In International Conference on Learning Representations. https://proceedings.iclr.cc/paper_files/paper/2025/hash/bbbd6d915cb90be21c1254a82d45cedd-Abstract-Conference.html.

[11] Schirrmeister et al. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping, 38 (11): 5391–5420. https://doi.org/10.1002/hbm.23730.

[12] Kastrati et al. (2025). EEG-Bench: A benchmark for EEG foundation models in clinical applications. https://arxiv.org/abs/2512.08959.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for the comments and discussion. We have uploaded the revised manuscript, with changes marked in blue.

The revision includes age-matched diagnosis results (Appendix D), patient-level bootstrap 95% confidence intervals (Appendix F), details of qEEG artifact rejection and retained signal duration (Section 4.3 and Appendix B), metadata field descriptions and diagnostic annotation procedures (Appendix A), and BASED score distributions and inter-rater reliability (Appendix E). We also added a quantitative result to the abstract and a Limitations and Future Directions section (Section 7).

As the first open EEG dataset dedicated to IESS, IESSEEG makes IESS EEG recordings, clinical metadata and clinician annotations available to the broader research community. We hope our data, benchmarks and analyses will help the clinical EEG and machine learning communities work together to investigate IESS and develop more effective methods for diagnosis and treatment-response prediction.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.