Learning-assisted transient-response analysis of nitride high-electron-mobility transistor aptasensors for rapid carcinoembryonic antigen detection.

Carcinoembryonic antigen (CEA) is widely used for tumor monitoring, yet electrical biosensor readouts are commonly based on endpoint or steady-state signals and therefore discard information contained in the early transient response. We developed a learning-assisted readout for an existing AlGaN/GaN high-electron-mobility transistor (HEMT) aptasensor by integrating transient drain-source current sequences with four steady-state I-V descriptors in a bidirectional long short-term memory-multilayer perceptron (LSTM-MLP) model. Bias-voltage comparison identified V_DS = 0.5 V as the preferred operating point on the basis of response amplitude, noise, and signal-to-noise ratio. Eighteen independent response curves spanning six experimentally tested CEA concentrations were evaluated by grouped three-fold cross-validation at the original-curve level, yielding R2 = 0.998, RMSE(log10) = 0.078, and MAE(log10) = 0.062. Curve-level low-concentration measurements analyzed with a fixed decision threshold and logistic detection-probability model produced a model-assisted C95 of 3.20 pg/mL (95% bootstrap confidence interval: 2.83-3.52 pg/mL). Sequence-length analysis showed that 120 s was the earliest interval meeting prespecified performance-retention criteria relative to the complete 1000 s record. In 1:10 diluted pooled human serum, baseline-corrected recoveries were 97.6%, 98.0%, and 95.0% at 5, 500, and 50,000 pg/mL, respectively; comparison with 1:20 serum further demonstrated dilution-dependent matrix effects. These findings show that transient-response learning can extract quantitative information not fully used by steady-state analysis, while broader device-batch and clinical validation remains necessary.