Data fusion of high-performance liquid chromatography and near-infrared spectroscopy with explainable machine learning for commercial origin authentication of Cimicifugae Rhizoma.

Multi-instrument data fusion combined with explainable machine learning offers a powerful strategy for complex sample authentication, yet quantitative attribution of each analytical source's contribution remains an unresolved methodological challenge. This study presents a two-platform data fusion framework integrating high-performance liquid chromatography with ultraviolet detection (providing both untargeted chromatographic fingerprints and targeted phenolic acid markers) and near-infrared spectroscopy, and employs SHAP (SHapley Additive exPlanations) TreeExplainer for quantitative source-level attribution under a standard fixed 440/110 train/test hold-out with strict data separation. Using 550 Cimicifugae Rhizoma samples from 11 Chinese provinces, we evaluated feature-level (dimensionally reduced concatenation), mid-level (43 features reduced by principal component analysis, PCA), and high-level (out-of-fold stacking) fusion strategies with Random Forest and XGBoost classifiers. Mid-level fusion with Random Forest achieved numerically perfect classification (0/110 errors; 95% CI 96.7-100%), but McNemar's test did not separate it from the best single-source baseline (2/110 errors, p = 0.500), whereas paired testing across 10 independent resampled splits did (mean difference +2.9 percentage points versus the chromatographic-fingerprint baseline, p = 0.0003). SHAP analysis quantified source contributions as chromatographic fingerprint 39.7%, targeted phenolic markers 26.2%, and near-infrared spectrum 34.0% (aggregating to 66.0% and 34.0% by platform); platform attribution was stable across 26 independent perturbations of seed, classifier, PCA dimensionality and training split (HPLC 65.6 ± 2.7%). The three targeted markers, constituting only 7.0% of features, contributed 26.2% of total importance (3.8-fold enrichment), with caffeic acid ranked first among all 43 features. PCA loading stability across data partitions was quantitatively confirmed via Hungarian algorithm matching (top-5 chromatographic principal components (PCs) mean |r| = 0.970; near-infrared PCs mean |r| > 0.999). PC loading back-mapping was less conclusive than the attribution itself: the highest-ranked chromatographic component loads on an unassigned early-eluting region rather than on the three targeted markers, and the near-infrared components carrying importance correspond to broad spectral-shape, O-H and C-H/O-H overtone variation rather than to phenolic-acid-specific bands; the cross-platform convergence supported by these data is therefore at the level of a bulk phenolic and matrix gradient, not of independent spectroscopic confirmation of the chromatographically identified markers. Because the three pharmacopoeial species of Cimicifugae Rhizoma segregate geographically and individual batches were not species-identified, species identity and province of collection are confounded; the discrimination modeled here is therefore a species-origin composite signal associated with province of collection, corresponding to the commercial consignments in which this multi-origin herbal drug is traded, and does not constitute proof of geographic causality. A systematic cross-comparison with six prior data fusion studies for botanical authentication contextualizes the methodological contributions, which are transferable to other multi-instrument analytical workflows.