Robust detection of femtogram-level Alzheimer's biomarkers using machine learning-enhanced graphene biosensors.

Early diagnosis of Alzheimer's disease (AD) requires blood biomarker tests sensitive to femtogram/mL concentrations. Graphene field-effect transistors (GFETs) are promising for this application, but suffer from device-to-device variability and require recalibration after functionalization. Here, we demonstrate a machine learning approach that overcomes these limitations, enabling robust AD biomarker detection without individual device calibration. By training artificial neural networks (ANNs) on full GFET transfer characteristics, our method automatically extracts features resilient to device variations. We detected three AD biomarkers-Aβ₄₂, Aβ₄₀, and P-tau217-at concentrations from 1 fg/mL to 1.0 × 10⁵ fg/mL with 98.9-100 % accuracy across multiple devices. Validation using 72 clinical plasma samples achieved four-way classification of cognitive states (healthy control, subjective cognitive decline, mild cognitive impairment, and AD), with multi-biomarker combinations improving diagnostic performance. SHAP (SHapley Additive exPlanations) analysis revealed that ANNs exploit previously uncharacterized regions of the GFET transfer characteristics that are not captured by conventional figures of merit. Unlike traditional approaches requiring device-specific calibration curves, our platform enables sensor deployment and maintains performance despite fabrication inconsistencies. This work demonstrates that machine learning can transform inherently variable graphene biosensors into reliable diagnostics, addressing a critical barrier to their potential implementation in point-of-care AD screening.