Dev2506/PQFL_Schizophrenia

PQFL Schizophrenia — Multi-Cohort Functional Connectivity Benchmark

Anonymized, per-subject 100×100 functional-connectivity/covariance matrices from four public schizophrenia-cohort repositories (N=936; 423 SZ, 288 HC, 225 BP), assembled for evaluating Personalized Quantum Federated Learning (PQFL) and baselines under a strict leave-one-domain-out (LODO) protocol. This repository accompanies the paper "Evaluating Personalized Quantum Federated Learning for Multi-Cohort Schizophrenia Classification: A Comparative Study" and supersedes earlier (retracted) results — see Correction notice below.

Cohort composition

Domain	n	SZ	HC	BP	Scanner	Source
LA5c (UCLA CoNPhOs)	189	50	122	17	Siemens Trio 3T, fMRIPrep derivatives	OpenNeuro ds000030
COBRE	146	72	74	0	Siemens Trio 3T, NIAK preprocessed (Figshare release)	SchizConnect / COBRE
TCP (Transdiagnostic Connectome Project)	130	13	92	25	Siemens Prisma 3T, two U.S. sites (Yale, McLean)	TCP release
Kaggle Psychosis	471	288	0	183	mixed 3T, group-ICA FNC features	Kaggle competition release
Total	936	423	288	225	4 repositories

The binary SZ-vs-HC screening subset (any domain lacking both classes is excluded; Kaggle contains no HC) is N=423 across 3 folds: LA5c (172), COBRE (146), TCP (105).

Representation caveat (important)

The common 100×100 matrix shape does not imply a common feature basis. LA5c, COBRE, and TCP provide ROI time series (Pearson FC over the Schaefer-100 parcellation); the Kaggle repository distributes group-ICA FNC vectors, reconstructed to a 105×105 matrix and truncated to the first 100 ICN rows/columns. Kaggle therefore occupies a different feature basis than the other three cohorts — a documented limitation, not a claim of a common representation.

Contents of this repository

• pqfl_project/ — experiment code, per-fold definitions, and result artifacts (analysis results, CV results, figures, final results from successive evaluation tiers)
• fig*.png (root) — legacy figures from the initial (superseded) project stage
• manuscript/ — the compiled manuscript PDF and source (final, corrected version)
• manifest*.csv, data_inventory.txt, site_effect_analysis.txt — dataset provenance records

Subject-level neuroimaging arrays are not redistributed here; they remain available from their original public repositories (OpenNeuro, SchizConnect/COBRE, TCP, Kaggle). The code and result artifacts in this repository are the research outputs; the underlying data follow their original licenses (see DATA_LICENSES.md).

Method summary

• Riemannian preprocessing: subject covariance/FC matrices regularized (C + λI, λ=1e-3) to SPD; Log-Euclidean tangent projection at the per-fold Fréchet mean; PCA to d=71 components (fit on outer-training subjects only).
• Harmonization: per-fold ComBat with site as batch and diagnosis as preserved covariate, estimated strictly on outer-training domains; held-out domains are left unadjusted.
• Model: 6-qubit HybridVQC — shared classical encoder 71→35→12→6, parallel 71→128 classical skip path, concat(130) → personal classifier head — with a FedPer shared/personal split (12,632 shared / 8,713 personal for C=3).
• Protocol: 10 seeds × 4 LODO folds per condition, 9 learning conditions (7 federated + centralized pooling + isolated local training) and 13 baseline models; fold-clustered statistics (n=4; Wilcoxon p-floor 0.125).
• Simulation: idealized statevector simulation throughout (no quantum-hardware noise).

Headline findings (corrected evaluation)

1. Centralized pooling attains the highest observed mean AUC (0.583).
2. PQFL does not win: at AUC 0.519 it is statistically indistinguishable from isolated local training; FedAvg (0.541) and FedBN (0.545) rank higher among FL methods.
3. Classical baselines lead overall: best is the polynomial SVM (AUC 0.611).
4. Deep ensembling (mean probability over 10 seeds) lifts the pooled binary PQFL from AUC 0.639 to 0.660 vs SVM 0.673.
5. No balanced-accuracy comparison survives Holm correction over the 21-test family (Wilcoxon p-floor at n=4 is 0.125); an exploratory 100-permutation null (configuration- mismatched) saw zero permutations reach the observed values.

Correction notice

A prior version of this work reported a significant PQFL advantage over an SVM baseline (AUC 0.597 vs 0.567, p=0.048) on a 4-class task, plus a large binary-screening advantage (d=0.85). Both claims are retracted:

1. The earlier "federated" configuration trained a single pooled client (no server, no aggregation, no multi-client loop). The earlier 3-class AUC 0.584 corresponds to pooled training, not federation (genuine FedPer federation yields 0.519).
2. The earlier binary comparison used the Kaggle domain as a held-out fold; that domain contains zero healthy controls, making balanced accuracy degenerate and AUC undefined there (sklearn returns 0.500 by convention). The corrected protocol excludes any domain collapsing to one class in a given scheme.

Figures and result files in this repository dated before 2026-08 reflect the superseded pipeline. The manuscript/ directory contains the corrected, final evaluation.

Limitations

• Only four evaluation domains; fold-level tests cannot reach p<0.05, and overlapping LOSO/LODO training sets induce dependence — all comparisons are descriptive.
• Kaggle's ICN basis differs from the Schaefer ROI basis of the other cohorts.
• Preprocessing statistics (Fréchet mean, PCA, ComBat) are pooled across training domains — a centralized coordination step, not itself federated.
• Explainability attributions come from a simplified 8-qubit surrogate (BA 0.517), not the deployed 6-qubit model.
• Results are from statevector simulation, not NISQ hardware.

Citation

@dataset{bansal_pqfl_schizophrenia,
  author  = {Bansal, Dev Datya Pratap and Puri, Krishna},
  title   = {PQFL Schizophrenia: Multi-Cohort Functional Connectivity Benchmark},
  year    = {2026},
  url     = {https://huggingface.co/datasets/Dev2506/PQFL_Schizophrenia}
}

Ethics

All data are de-identified secondary data from public repositories. Primary collection was approved by the respective institutional review boards with written informed consent. This secondary analysis follows the documentation supplied by the original repositories.