Accurate monitoring in complex biofluids remains challenging due to unpredictable skin - electrode interfacial interference and matrix-dependent impedance variations. Here, we present an intelligent enzymatic microneedle platform that achieves calibration-efficient, matrix-adaptive, and drift-resilient high-fidelity sensing through a Nafion-shielded dual-mode architecture and an explainable feedforward neural network (FFNN). Using glucose as a representative model analyte, the platform integrates 2T/4T impedance acquisition with a functionalized Nafion interface to improve robustness against contact-related and biofluid-induced perturbations. SHapley Additive exPlanations were used to interpret the internal decision logic of the FFNN, revealing that the model learns physics-aligned feature relationships rather than relying solely on statistical fitting. Specifically, the network identifies matrix-dependent compensation patterns associated with contact impedance and leverages phase - impedance coupling to evaluate interfacial integrity before processing glucose-related electrochemical responses. This interpretability analysis suggests that the model can distinguish glucose-associated enzymatic signals from nonspecific environmental and interfacial perturbations. After few-shot domain adaptation, the platform achieved a mean absolute relative difference (MARD) of 2.55% in whole blood, demonstrating high prediction fidelity in a complex biological matrix. Supported by ablation studies and principal component analysis, this work bridges electrochemical interface engineering and explainable machine learning, offering a highly scalable and robust framework for minimally invasive metabolic monitoring.
Piao et al. (Tue,) studied this question.