Process-aware inference of biomedical Raman spectra classification

The evaluation of classification models is mostly based on accuracy. However, biomedical data often contain heterogeneous, spatially or spectrally distributed contributions that may provide partially conflicting evidence during classification. In such settings, analyzing how this evidence is reconciled during inference can complement conventional performance metrics and provide additional insight into model behavior. Here, we introduce a process-aware inference framework in which local evidence sources interact through regulated negotiation, generating a measurable inference trajectory. We instantiated this framework in Raman spectroscopy by partitioning spectra into adjacent macro-bands that function as semi-autonomous agents contributing local class evidence. From the trajectories, we defined Decision Geometry (DG) as a descriptor of total reconciliation effort and effective Regulated Early Decision Geometry (eREDG) as a descriptor of early-stage negotiated dynamics. We evaluated the framework on a three-class melanoma Raman dataset, a small grouped in vivo diabetes Raman cohort, and a larger external 5-class bacteria/yeast Raman dataset comprising 10,000 training/validation spectra and 2500 independent external-test spectra. We showed that conventional classification accuracy cannot capture qualitative differences in inference trajectories, whereas DG reveals spectral conflict and regulatory strength. Under regulated regimes, eREDG further stratifies correct and incorrect predictions without access to ground truth. Finally, band-resolved H1 conflict analysis was localized to Raman regions. We propose this framework as a proof-of-principle, process-aware method to complement conventional machine-learning pipelines, demonstrated here in Raman spectroscopy as a domain where inference depends on spectral bands evidence.