Ultra-early detection of S100B biomarkers using a nanophotonic biosensor with deep learning quantification: A clinical model based on EDAS patients.

BACKGROUND

Ultra-early detection of brain injury biomarkers within the critical first hour post-injury remains a major clinical challenge in mild traumatic brain injury (mTBI) management. Conventional platforms (e.g., ELISA) suffer from limited sensitivity, prolonged turnaround times, and narrow dynamic range, hindering timely diagnosis.

METHODS

We developed an integrated nanophotonic biosensor and deep learning platform for ultrasensitive S100B quantification. The biosensor employs capillary-assembled heterochain arrays of 500-nm polystyrene and 250-nm selenium nanoparticles, enabling red-channel optical amplification. A dedicated deep learning framework (ProSpect), incorporating ResNet-50 backbone and probabilistic patch aggregation, was engineered for automated image analysis. Encephaloduroarteriosynangiosis (EDAS) patients (n = 25) were leveraged as a time-controlled human model of mild cortical injury. Matched biospecimens (serum, urine, saliva, CSF) were collected at preoperative, immediate postoperative (≤15 min), and 1-h postoperative time points for parallel biosensor and ELISA analysis.

RESULTS

The biosensor achieved a detection limit of 1 pg/mL, a dynamic range spanning 1 pg/mL-100 ng/mL (R2 = 0.9995), and intra-assay CV of 4.21 %. ProSpect-based quantification attained 98.08 % accuracy in semi-quantitative classification and near-perfect ROC-AUC (0.999). Clinically, 47.6 % of samples were undetectable by ELISA (LOD = 18.75 pg/mL) but successfully quantified by the biosensor. Strong inter-method correlation was observed (R2 = 0.9884). Significant postoperative S100B elevations occurred in serum, while salivary levels decreased and urinary concentrations remained stable.

CONCLUSION

This biosensor-AI platform enables rapid (<30 min), matrix-flexible, and ultrasensitive S100B detection, overcoming limitations of conventional assays. Validated through the EDAS clinical model, it holds significant potential for point-of-care deployment in mTBI screening, neurotrauma triage, and real-time neurocritical monitoring.