This study compares quantum machine‑learning models with classical counterparts for classifying neuropsychiatric disorders using resting‑state functional connectivity (rs‑FC) data. Four public cohorts—ABIDE I, ADHD‑200, REST‑meta‑MDD, and UCLA‑CNP—comprising 4,308 subjects were used, and all models were fed identical correlation‑derived features.
Leave‑site‑out validation was applied to ABIDE I, ADHD‑200, and REST‑meta‑MDD, while stratified 10‑fold cross‑validation was used for UCLA‑CNP. Six families of quantum models (quantum kernel, trainable quantum kernel, hybrid quantum convolutional network, etc.) were benchmarked against matched classical kernels or convolutional networks.
The quantum kernel was statistically indistinguishable from the matched classical kernel across all four cohorts, with the average accuracy gap staying within 1.4 percentage points at a 95 % confidence level. On ABIDE I, a hybrid quantum convolutional network achieved 0.645 accuracy versus 0.641 for its classical counterpart, and a trainable quantum kernel improved AUC by 0.064 over the classical kernel. Quantum convolutional networks also yielded higher AUC on this cohort and exhibited smaller train‑test gaps across all datasets, though without a clear accuracy gain.
Overall, most quantum models matched their classical equivalents, and none consistently outperformed them. This work provides a feature‑matched benchmark of 284 quantum‑classical comparisons and represents one of the first quantum evaluations of ABIDE I under leave‑site‑out validation.
Additional analyses of fidelity‑based graph construction, effective‑dimension, and eigenspectrum were performed to account for the observed quantum‑classical parity.
Review