Decoupling human hair isotope codes via mechanistic feature construction and calibrated evidence fusion for...

Stable isotope analysis of human hair (δ²H, δ¹⁸O, δ¹³C, and δ¹⁵N) offers a non-invasive trajectory for geographic sourcing, yet its forensic application remains constrained by overlapping isotopic catchments, individual lifestyle variations, and the high risk of overinterpretation in deterministic machine-learning classifiers. Here, we present a chemically interpretable and uncertainty-aware framework that transitions hair isotope forensics from empirical pattern recognition to rigorous evidence quantification. To impose a chemistry-informed representation on the correlated isotope measurements, we constructed the Resident-Hair Water-Diet Isotope Descriptor (RH-WDID) system by integrating hydrological gradients, dietary contrasts, and cross-domain coupling features. Crucially, all mathematical transformations were constrained strictly within training folds to eliminate information leakage. Compared with the raw isotope space, the RH-WDID feature space markedly enhanced spatial compactification and separability, increasing the unsupervised silhouette coefficient from 0.178 to 0.305 and reducing the Davies-Bouldin index from 1.517 to 1.325. Systematic benchmarking identified a residual tabular multilayer perceptron (ResidualTabMLP) as the optimal discrimination backbone, achieving a high internal five-fold cross-validated accuracy of 0.937, with mechanistic validity supported by SHAP analysis and forward/reverse ablation experiments. Importantly, this value represents internal cross-validation performance only; no independent external test cohort collected from another region, collection period, or laboratory was available in the present study. Therefore, the model should be interpreted as a first-stage, scope-defined seven-city resident-reference framework rather than as a fully externally validated forensic classifier. To reduce the risk of overinterpretation, this deterministic backbone was further upgraded into a Resident-Hair Forensic Isotope Evidence-Fusion (RH-FCIEF) model. Instead of forced-choice top-1 labels, RH-FCIEF outputs a calibrated multi-model probability vector coupled with scope-aware typicality assessments. This study provides a chemically interpretable and uncertainty-aware modelling route for converting bulk isotope measurements into probabilistic resident-source evidence within a clearly defined reference scope.