Explainable Rhythm-Resolved EEG for Retrospective Comparison of Psychiatric-Service Adults with Recorded Violent Offences and Non-Clinical Controls

Background/Objectives: This study examines whether rhythm-resolved EEG features separate two retrospective groups and whether those features can be traced to explicit rhythms and sensors. The groups are psychiatric-service adults with recorded violent offences and non-clinical controls; the design does not compare offenders with non-offenders. Methods: We analysed 2554 non-overlapping 15 s EEG epochs from 95 adults (51 controls; 44 recorded-offence participants). RhythmFluxPat generated 9216 rank-transition features. CWImRMR selected 352 features with full-cohort labels, and classifier selection was nested within 95 participant-held-out leave-one-subject-out (LOSO) folds. Results: Classifier-nested tkNN classified 2392/2554 epochs (93.66%) and 93/95 participants (97.89%), and an independent check reproduced both confusion matrices. A nested support-vector machine (SVM) gave similar values (94.71% of epochs; 93/95 participants). In the comparative evaluation of four feature selectors, CWImRMR remained the best branch (93.85%), and a 44-feature subset reached 91.11%. DRLobish linked the selected features to delta and beta-to-gamma routes in the recorded-offence-assigned panel and to theta and alpha-to-theta routes in the control-assigned panel. Conclusions: RhythmFluxPat provides traceable rank-persistence and rank-exchange coordinates. The reported accuracies are optimistic internal estimates. Fully nested replication with matched psychiatric controls is required before any claim of generalisation or biomarker validity.