keremkem/eegdirectsignalvoxelization: v5.1.0 - Frequency- Specific with Automated LOSO & Weights
Abstract
[v5.1.0] — 2025 · "Frequency-Specific" (LEMON Dataset · Eyes-Open) New Features Relative band power (hilbert_envelope_relative): Replaces absolute Hilbert envelope. Normalizes each band's power by total broadband power, eliminating 1/f covariance artifacts that previously inflated cross-band correlations. Rank-based Dice similarity: Amplitude-invariant voxel overlap scoring. Eliminates sensitivity to broadband power fluctuations; particularly beneficial for gamma and delta bands where absolute amplitudes vary widely across subjects. CLI argument parser (argparse): Pipeline can now be invoked from the command line with flags (e.g., --no-ar1, --no-loso, subject filtering). Per-subject QC CSV export: Quality-control metrics written to structured .csv files alongside NIfTI outputs for downstream statistical review. Thread-safe Logger (threading.RLock): Logger class upgraded with reentrant lock for safe parallel logging across LOSO folds. Bug Fixes Gamma spatial prior corrected: Was mistakenly inheriting the alpha spatial prior (center_y = -75). Now uses a frequency-appropriate dual-peak prior (posterior visual + frontal cognitive). Affects all gamma-band voxel maps. HRF variance scaling removed: Post-convolution variance normalization was incorrectly applied inside the pipeline. CONN/SPM handle normalization internally; double-scaling has been removed. Output NIfTI volumes are now fully CONN/SPM/FreeSurfer-compatible. HRF mmap edge padding added: Memory-mapped HRF convolution path was missing edge padding, causing boundary artifacts in chunked z-slice processing. Fixed. Changes LOSO targets recalibrated: All literature-derived validation targets (posterior/anterior alpha ratio, sensorimotor beta, frontal-midline theta, gamma posterior emphasis, anterior/posterior delta) recalibrated for the relative power scale. Absolute-power targets from v3.x/v4.x are no longer applicable. Dataset: SPIS (N=10, eyes-closed) → LEMON (N=40, eyes-open). Subject list updated accordingly (sub-03230x_EO.mat format). Sampling rate: 256 Hz → 250 Hz (LEMON dataset hardware). Cache version bumped to v51: Incompatible with cache_v40 and cache_v50. Old caches must be regenerated. Additional imports: csv, argparse, threading, zscore, rankdata, gaussian_filter, interp1d, LinearRegression, shutil added to support new features. Codebase size: ~2,890 → ~3,483 lines (+593 lines).
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 4 March 2026