Person‐Centric Attribution From Genus‐Level Microbiome Compositions: A Leakage‐Controlled Machine‐Learning Pipeline With Feature‐Stability Analysis and Permutation‐Test Assessment of Timepoint Signal

This study evaluated whether genus‐level microbiome profiles obtained from five individuals at four timepoints ( N  = 20 samples and D  = 323 genera) contain a reproducible signal for PERSON attribution and, secondarily, TIME classification in a forensic setting. PERSON attribution (5‐way) was assessed under leave‐one‐time‐out (LOTO) validation to test whether person‐associated patterns remain detectable at unseen timepoints. TIME classification (4‐way) was assessed under leave‐one‐person‐out (LOPO) validation to examine whether any temporal signal transfers to unseen individuals. The analysis combined centered log‐ratio preprocessing, training‐fold feature selection with ReliefF and mRMR, classification with RBF‐SVM, kNN, random forest, and AdaBoostM2, and probability‐level ensembling. Under LOTO, PERSON attribution reached macro‐F1 = 0.60 and macro‐AUC = 0.81. Under LOPO, TIME prediction remained near chance (macro‐AUC = 0.51), and permutation testing did not support a generalizable time‐related signal ( p = 0.42). Bootstrap confidence intervals, confusion matrices, learning curves, and feature‐stability analyses are reported to support a cautious interpretation of these findings.