Paired-reference full-spectrum modeling with sensor-chip-grouped validation for LSPR quantification of...

Localized surface plasmon resonance (LSPR) assays commonly use peak shifts for quantification, discarding line-shape information. Randomly splitting correlated spectra from the same sensor chip may also overestimate model performance. We developed an annealing-induced gold nanoparticle/fluorine-doped tin oxide (AuNP/FTO) immunosensor and a paired-reference full-spectrum framework for carcinoembryonic antigen (CEA) estimation using sensor-chip-grouped nested validation. After quality control, 440 paired bovine serum albumin (BSA)-reference/CEA-response spectra from 32 independent chips covering 11 concentrations across a modeling range of 0.5-100 ng/mL were analyzed. Peak shifts were approximately linear from 0.5 to 10 ng/mL (R² = 0.9695), with a calculated limit of detection of 0.274 ng/mL. Among seven prespecified regressors, a fusion convolutional neural network (Fusion CNN) combining response spectra, first derivatives, and BSA-reference descriptors achieved the lowest concentration-balanced chip-level root-mean-square error on the log10-transformed concentration scale (RMSE_log10 = 0.126 ± 0.028). Zeroing or mismatching the reference descriptors significantly increased error, supporting the contribution of correct reference pairing. No statistically significant difference was detected between difference-spectrum partial least-squares regression (PLSR) and Fusion CNN. Preliminary selectivity testing and a single-patient serum-dilution experiment provided initial support for analytical feasibility. This study provides a paired-reference LSPR quantification framework that combines full-spectrum representation with leakage-resistant chip-level validation.