Psychopathology prediction

Name: psychopathology

Category: Phenotyping

Dataset: Shirazi2024 (HBN)

Objective: Regression

Split: Leave-subjects-out

Usage

neuralbench eeg psychopathology

Show config.yaml

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.

data:
  study:
    source:
      name: Shirazi2024Hbn
    filter_resting_state_with_externalizing:
      name: QueryEvents
      query: "(type == 'Eeg') & (task == 'task-RestingState') & (duration > 180.0) & externalizing.notnull()"
    crop_timelines:
      name: CropTimelines
      event_type: Eeg
      start_offset_s: 60.0
      max_duration_s: 120.0
    split:
      name: PredefinedSplit
      test_split_query: "release in ['R5']"
      col_name: split
      valid_split_by: release
      valid_split_ratio: 0.091  # 1/11
      valid_random_state: 33
  target:
    =replace=: true
    name: EventField
    event_types: Eeg
    event_field: externalizing
    aggregation: single
  trigger_event_type: Eeg
  start: 0.0
  duration: 2.0
  stride: 2.0
  summary_columns: [release, externalizing]
brain_model_output_size: &brain_model_output_size 1
trainer_config:
  monitor: val/pearsonr
  mode: max
  strategy: auto
  patience: 7
  n_epochs: 40
loss:
  name: MSELoss
metrics: !!python/object/apply:neuralbench.defaults.metrics.get_regression_metric_configs
  - *brain_model_output_size

Description

This task corresponds to Challenge 2 (“Externalizing Factor Prediction / Subject-Invariant Representation”) of the EEG Challenge 2025 [Aristimunha2025]. The goal is to predict the externalizing psychopathology factor – a continuous score derived from the Child Behavior Checklist (CBCL) – from 2-s resting-state EEG windows. This probes whether models can learn generalizable neural representations that transfer across subjects while recovering clinically relevant variability.

The official challenge metric is the normalized root-mean-squared error (RMSE divided by the standard deviation of the ground-truth externalizing scores), reported as normalized_rmse alongside the other regression metrics.

Dataset Notes

• Shirazi2024 (HBN) contains EEG recordings from 11 cohorts (“releases”) containing different participants. Here, we leave one release out for testing.

• The dataset contains different tasks (resting-state, contrast change detection, etc.). Here, we only use the resting-state data for psychopathology prediction.

• Only subjects with a non-null externalizing CBCL score are retained.

References

[Aristimunha2025]

Aristimunha, Bruno, et al. “EEG Foundation Challenge: From Cross-Task to Cross-Subject EEG Decoding.” arXiv preprint arXiv:2506.19141 (2025).