Robustness analysis of simulated biochip signal-based salmonella serovar classification under noise and sensor drift perturbations

Antimicrobial resistance (AMR) has become a major global healthcare challenge, necessitating rapid, accurate, and reliable diagnostic systems capable of operating under realistic clinical and biosensor conditions. Although machine learning-assisted biosensor platforms have shown promising predictive performance, the robustness of these systems under signal noise and sensor drift perturbations remains insufficiently explored. This study aimed to develop and evaluate a robustness-oriented computational framework for Salmonella serovar classification using AMR-associated genomic features represented through simulated biochip signals integrated with ensemble machine learning and explainable artificial intelligence (XAI). The framework specifically investigated predictive stability, interpretability robustness, and generalization behavior under progressive signal degradation conditions. A publicly available pathogen detection dataset containing over 400,000 genomic records was utilized, from which 10,000 stratified samples with 213 antimicrobial resistance genes were selected. Binary AMR genotypes were transformed into continuous-valued analog biochip signals using Gaussian-distributed signal encoding to emulate realistic biosensor outputs. A clean baseline condition (σ = 0.00) and progressive Gaussian noise perturbations (σ = 0.03–0.10), together with multiplicative sensor drift (δ = 0.02–0.08), were introduced to simulate environmental interference, calibration instability, and probe degradation. Four machine learning classifiers, including Random Forest, Support Vector Classifier (SVC), Voting Ensemble, and Stacking Ensemble, were evaluated using stratified 5-fold cross-validation. Model interpretability robustness was investigated using SHAP (SHapley Additive exPlanations) analysis. The Voting Ensemble demonstrated comparatively greater robustness and maintained relatively stable predictive performance across the investigated simulated perturbation scenarios. Under severe Gaussian noise (σ = 0.10), the model maintained an F1-score of 0.967 and AUC of 0.982, corresponding to only a 2.3% reduction relative to baseline performance. In contrast, SVC exhibited the greatest sensitivity to perturbation, with the F1-score decreasing from 0.972 to 0.921. Similarly, under severe sensor drift (δ = 0.08), the Voting Ensemble preserved stable classification performance with an F1-score of 0.971. SHAP analysis revealed that dominant resistance-associated genes, including APH(3’’)-IB, APH(6)-ID, SUL1, and MDSB, retained stable feature importance rankings despite substantial signal corruption. Learning curve analysis additionally demonstrated limited evidence of severe overfitting and strong generalization capability under noisy sensing conditions. The proposed noise and drift evaluation framework demonstrated that ensemble machine learning combined with robust explainable AI can improve the reliability, stability, and practical applicability of intelligent biochip-assisted AMR diagnostics. The findings establish a computational foundation for future experimental biosensor validation and robustness-aware AI-assisted diagnostic research.