Decoding mixed-toxicity via multimodal bioelectrochemical sensing: A cascaded framework for rapid assessment of...

Bioelectrochemical sensors often exhibit non-additive responses in mixed-toxicity matrices and, more importantly, lack a quantitative linkage between fast electrode-level signals and downstream process performance. Here, we develop a cascaded prediction framework that uses a S. oneidensis MR-1 electroactive biofilm sensor to translate rapid bioelectrochemical perturbations into activated-sludge inhibition indicators. A multimodal response feature space was constructed from cyclic voltammetry (CV), differential pulse voltammetry (DPV), and chronoamperometry (I-t) to characterize mixture-dependent inhibition behaviors induced by Hg²⁺, tetrahydrofuran (THF), and formaldehyde (HCHO). Using mechanism-inspired but phenomenological feature transformations and full-spectrum descriptors, feature-enhanced XGBoost and LightGBM models enabled mixture-aware quantification of the three toxicants (average R² = 0.82 on a held-out test split). In particular, HCHO prediction improved from R² = 0.653 to 0.840 after incorporating kinetic-focused descriptors. The estimated toxicant concentrations were then used as intermediate variables in a second-stage mapping to predict activated-sludge inhibition quantified by the specific oxygen uptake rate (SOUR) index, achieving R² ≈ 0.898. Robustness of the cascaded framework to upstream prediction uncertainty was evaluated via Monte Carlo error propagation using a nearest‑neighbor wild bootstrap with the Mammen distribution (k = 5). The sensor response remained stable against common matrix variations (pH 5-7 and salinity up to 30 g·L⁻¹) under anoxic operation. Overall, the framework provides a proof-of-concept route for linking rapid bioelectrochemical responses to process-relevant inhibition indicators within the calibrated operating domain.