Abstract
Resting-state functional connectomes are widely used to discriminate patients with schizophrenia from healthy controls, yet reported cross-validated accuracy depends heavily on the point at which features are defined relative to the train/test split. We audit this dependence on 146 participants of the COBRE cohort parcellated with the Schaefer-100 atlas (4,950 edges). Exploratory covariate-adjusted mass-univariate inference is kept strictly separate from prediction, and every predictive pipeline performs edge screening, dimensionality reduction and hyperparameter search inside the outer training folds of a repeated nested cross-validation. Under this regime, elastic-net logistic regression reaches an AUC of 0.768 ± 0.076 and PCA-30 an AUC of 0.771 ± 0.079, with an edges-only elastic net essentially unchanged at 0.765 ± 0.079. Fold-wise residualisation of mean framewise displacement lowers that edges-only estimate to 0.694. Holding a twenty-edge |t| screen and a logistic classifier fixed, and varying only whether the screen sees the test fold, isolates leakage for that construction: AUC rises from 0.709 ± 0.076 to 0.903 ± 0.039. That 0.194 gap is the cost of leaking a hard univariate screen, not of leaking the primary elastic net. The same construction on a 50/50 age-, sex- and FD-matched subsample of UCLA CNP yields 0.679 against 0.873, again a 0.194 gap; that analysis is within-cohort nested cross-validation, not a locked transfer, and does not revise 0.77. On synthetic data with no signal the leakage-free protocol returns chance, whereas a leaked twenty-edge screen reports 0.82 when the edges are correlated as in a connectome and 0.94 when they are independent; the observed leaked value of 0.903 lies between those bounds and about 0.08 above the correlated floor. Calibration and subject-level aggregation of repeated out-of-fold predictions are reported for the leakage-free model. The resulting figure of roughly 0.77 AUC is specific to this cohort, parcellation, connectivity measure and model family rather than a general ceiling, and the audit protocol is released as public code.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MLMI_035.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=gA855MbpOC
BibTex
@InProceedings{ChaIsh_LeakageControlled_MICCAISAT2026,
author = { Chandok, Ishsirjan Kaur},
title = { { Leakage-Controlled Resting-State Connectome Classification of Schizophrenia: A Nested Cross-Validation Audit of the COBRE Cohort } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17272},
month = {pending},
page = {pending}
}