An AI-Ready Pipeline for Impedance-Resolved QCM Biosensor: Interpretable Line-Shape Features, Redundancy Control, and Robust Regression

Accurate inference from quartz crystal microbalance (QCM) measurements in liquids is often limited by reducing resonance behavior to two scalar endpoints (frequency and dissipation shifts, $\mathrm{\Delta }f$ and $\mathrm{\Delta }D$) or by relying on single-equation analytical models (Kanazawa-model). We propose an AI-ready, impedance-resolved workflow that preserves full resonance line-shape information and converts it into compact, physically interpretable features for supervised regression. A passive microfluidic mixer generates glycerol–water concentration gradients under a constant total flow rate (50~$\mu$L/min), while complex impedance spectra of a 10~MHz AT-cut quartz crystal are recorded in real time. Each sweep of nine spectra is parameterized via constrained Gaussian/Lorentzian models to yield 52 line-shape descriptors spanning extrema of $\mathrm{\mid }Z\mathrm{\mid }$, $X$, and $B$ and peaks of $R$, phase, and $G$. The pipeline integrates consensus outlier handling, redundancy-aware feature ranking (mRMR), and cross-validated regression across linear, kernel, and ensemble models. Compared with the classical Kanazawa baseline, the impedance line-shape approach reduces concentration-prediction error from 0.456 to 0.148~%v/v RMSE (3.09$\times$ lower). The results demonstrate that impedance-resolved line-shape features provide a robust and interpretable basis for machine-learning-assisted QCM inference and illustrate a generalizable pattern for AI-enabled spectral sensing.