SpecCal-Transfer: Physics-Guided Augmentation-Based Spectral Calibration Transfer for Reagent-Strip Urinalysis

Calibration transfer for portable reagent-strip spectroscopy is challenged by optical and instrumental variation. SpecCal-Transfer examines physics-guided spectral augmentation using 288 SpectroChip spectra representing 12 reagent-strip response classes, eight nominal control-mixture fractions, and three repeated acquisitions. Simulated gain, baseline drift, wavelength displacement, noise, spectral smoothing, missing-band corruption, and combined changes were used to evaluate source-only regression, shift-aware augmentation, neural correlation alignment (CORAL), maximum mean discrepancy (MMD), conventional standardization, and simulated limited-reference adaptation. In an extremely randomized trees (ExtraTrees) evaluation, physics-guided training reduced moderate wavelength-shift root-mean-square error (RMSE) from 0.1476 to 0.1233 (16.53% mean paired reduction) and combined-shift RMSE from 0.1504 to 0.1232 (18.17%). An independent training-size control yielded 0.1240 with expanded augmentation versus 0.1502 without augmentation under moderate wavelength shifts; size-matched gains were smaller. Under wavelength perturbations at previously withheld interior fractions, physics-guided augmentation achieved RMSE 0.1612 versus 0.1661 for a magnitude-matched generic control. Neural alignment and simulated recalibration had condition-dependent effects. These results establish computational robustness under specified measurement mechanisms, not experimental inter-device transfer or analyte-specific concentration accuracy.