A dual-band radio-frequency biosensing platform with physics-informed deep learning for simultaneous sweat pH and...

Continuous monitoring of sweat pH and lactate is of great value for physiological assessment, yet existing sensors face two critical bottlenecks: prolonged response times that hinder real-time tracking, and signal decoding relying on simple linear calibration that fails to handle nonlinearities and dynamic hysteresis. Here, we report a synergistic "RF-materials-AI" dual-band biosensing platform integrating (1) a stub-loaded dual-resonance microstrip antenna at 2.61 GHz and 3.98 GHz for frequency-domain discrimination of pH and lactate responses; (2) ternary nanocomposite interfaces-rGO/TiO₂/MWCNTs for pH and rGO/CuO/NiCo-MOF for lactate-for dielectric modulation and redox-mediated charge transfer, respectively; and (3) a Physics-BiLSTM-Attention network embedding first-order temporal difference priors and a dual-branch attention mechanism for robust multiparameter regression. The sensor achieves response times of 55 s (pH) and 135 s (lactate) with good stability and selectivity. On human sweat data from 10 subjects under LOSO-CV evaluation, the model delivers lactate prediction with R = 0.98, R²=0.96, MAE = 0.87 mM, RMSE = 1.12 mM, and pH prediction with R = 0.94, R²=0.88, MAE = 0.12 pH units, RMSE = 0.15 pH units-substantially outperforming static calibration (29.8% and 31.4% MAE reduction, respectively). The model achieves 94.0% accuracy in health status classification and 91.2% in exercise intensity classification, with well-calibrated uncertainty [Formula: see text]. Ablation studies confirm that removing any core component degrades MAE by 27-73%, underscoring the indispensable synergy of the three design dimensions. This work establishes a materials-hardware-algorithm co-design framework for multiparameter sweat analysis, offering a conceptually new pathway toward intelligent, passive, and continuous biosensing.