Edge-AI-enabled Aptamer Impedance Biosensor for Real-time Point-of-care Tacrolimus Therapeutic Drug Monitoring

Introduction: Tacrolimus therapeutic drug monitoring (TDM) is essential after kidney transplantation given its narrow therapeutic window (5–15 ng/mL) and pronounced inter- and intra-patient pharmacokinetic variability. Conventional TDM relies on laboratory-based immunoassay or mass spectrometry, introducing delays that impair timely dose adjustment. Point-of-care (POC) biosensors offer a promising alternative, but achieving sub-ng/mL accuracy in complex whole-blood matrices while maintaining clinical deployability remains an unresolved challenge. We developed an edge-AI-enabled aptamer-based electrochemical impedance spectroscopy (EIS) biosensor for real-time, privacy-preserving POC tacrolimus TDM and validated it prospectively in kidney transplant recipients.
Methods: A methylene blue–labeled tacrolimus-binding aptamer was immobilized on a flexible 3×3 gold microarray electrode. A two-point EIS measurement protocol—comprising a reference solution sweep followed by a whole-blood sweep—was employed to isolate analyte-induced differential signals and suppress sensor-to-sensor variability. Phase-based quality control was applied to exclude aberrant sweeps. Multiple complex-plane features were engineered from broadband EIS data, including phase difference, Euclidean distance, difference vector angle, dot and cross products, and impedance magnitude ratio. These feature sequences were processed by a sliding-window bidirectional long short-term memory (Bi-LSTM) network trained to learn cross-frequency spectral patterns, enabling lightweight on-device inference without data transmission. The system was evaluated using venous whole-blood samples from 87 kidney transplant recipients (206 EIS sweeps), with the Siemens ACMIA assay serving as the reference standard.
Results: On a held-out test set, the Bi-LSTM model achieved sub-ng/mL prediction accuracy (mean absolute error 0.39 ng/mL; root mean squared error 0.79 ng/mL). Bland–Altman analysis demonstrated strong agreement with the ACMIA reference method (mean bias −0.30 ng/mL; 95% limits of agreement −1.45 to 0.86 ng/mL), well within clinically acceptable margins for tacrolimus TDM.
Conclusion: An edge-AI-integrated aptamer-EIS biosensor achieves clinically actionable tacrolimus measurement accuracy directly from whole blood, without laboratory infrastructure. The system's privacy-preserving on-device inference architecture supports future integration into closed-loop TDM platforms. These findings establish a translational foundation for real-time, AI-guided precision immunosuppression management at the point of care.