AI-Driven Dual-Graphene Near-Infrared SPR Biosensor for Non-Invasive Salivary Cortisol Monitoring

Non-invasive salivary cortisol measurement requires separation of receptor binding from matrix effects and measurement artefacts. We combined a 1550 nanometre dual-graphene optical transducer with an antibody-functionalised cortisol assay and two separately evaluated artificial intelligence workflows. For structural modelling, 8192 scrambled Sobol proposals yielded 7941 accepted geometries. Grouped development selected an extremely randomised tree ensemble, whose sealed-test coefficient of determination was 0.986. Genetic and particle swarm searches were performed after model freeze, and direct transfer-matrix calculations gave a leading candidate sensitivity of 198.60 degrees per refractive-index unit and a resonance width of 0.1790 degrees. These computational candidates were distinct from the nominal cartridge used experimentally. The receptor-functionalised assay had a detection limit of 0.15 nanograms per millilitre and a validated quantitative range of 0.25–8.00 nanograms per millilitre. A gradient-boosted model using complete target and reference angular traces achieved mean absolute errors of 0.25 and 0.50 nanograms per millilitre in 25 blinded artificial-saliva samples and 40 human-saliva samples, respectively; conventional calibration gave corresponding errors of 0.34 and 0.62. Human-saliva measurements were compared with isotope-dilution liquid chromatography–tandem mass spectrometry. Nominal 90% prediction intervals achieved 92% and 90% coverage. The nominal optical advantage over the simpler AZO/graphene reference was incremental, and residual cortisone cross-reactivity remained. These results support a proof-of-concept biosensor requiring further manufacturing and external clinical evaluation.