Nondestructive Detection Methods For Bacterial Foodborne Pathogens Based On Hyperspectral Imaging,Electronic Nose And Data Fusion Integration

Bacterial foodborne pathogens(BFP)are one of the main sources of food safety problems.Rapid detection of BFP is of great importance.In this work,Electronic nose(E-nose)and hyperspectral imaging(HSI)together with chemometrics are applied for rapid foodborne pathogen detection and quantification.Model performance is improved by Metaheuristic optimization and wavelength selection algorithms.We proposed a new data fusion approach of HSI,E-nose,and Headspace-gas chromatography-ion mobility spectrometry(HS-GC-IMS)to provide a higher synergistic effect on pathogen classification and discrimination.Application of HSI and E-nose as a process analytical technology for monitoring ultrasound inactivation of pathogens is also investigated.The first part of the research deals with discrimination and classification of BFPs at the species and strain level grown on agar and broth cultures by electronic nose and hyperspectral imaging.Initially,the application of E-nose for classification and differentiation of BFPs using single colony bacteria in phosphate buffer solution was explored.Linear discriminant analysis(LDA)and support vector machine(SVM)models were established,and classification accuracies for training set were 90.3%,91%and 89.5%,90.6%for the prediction set respectively.Particle swarm optimization(PSO)method was employed to enhance model performance,and classification accuracies were improved remarkably,100%for the training set and 98.95%for prediction set.Likewise,we explored the feasibility of improving and visualizing HSI spectral preprocessing using pixel-wise analysis for better classification of BFP grown on agar plates.LDA and SVM classification models provided classification accuracies of 91.7%,83.9%for training and 90.7%,82.2%for prediction sets respectively.Also,the c and g of SVM parameters were optimized,employing PSO to improve classification of the spectral data.Results showed that the PSO-SVM model got the accuracy of 100%and 98.44%for training and prediction sets,respectively.Furthermore,wavelength selection algorithms were applied to reduce the number of wavelengths.Competitive adaptive weighted sampling-particle swarm optimization-support vector machine model achieved the highest classification accuracy of99.48%and 98.44%for training and prediction set,respectively.In second part of the work,data fusion strategy of E-nose,HSI and HS-GC-IMS datasets were proposed and results produce better inferences than a solitary technique in discriminating BFPs.Further experiments were conducted for the development of an optimal integrated system by fusing the feature vectors extracted from HSI,E-nose and HS-GC-IMS data.Pathogens grown on agar plates and inoculated in pork was investigated for improving pathogen detection and classification.E-nose and HSI fused datasets were applied for BFP classification and discrimination at the strain and species level on agar plates.Results indicated that the fusion model showed improved accuracies with LDA improving to 99.17%,and98.75%and k–nearest neighbor to 98.75%and 98.33%for training and prediction set respectively.Microbial volatile organic compounds analysis by E-nose and HS-GC-IMS for early detection of microbial food contamination was also studied.The discrimination ability of the models built on solely E-nose or HS-GC–IMS datasets were satisfactory for both broth-based and inoculated pork samples.However,the data fusion approach resulted in a more robust classification model with greater than 98%accuracy and better discrimination ability.Our results suggest that volatile organic compound-profiling by E-nose and HS-GC-IMS data fusion has the potential of increasing prediction accuracy.We illustrated for the first time data fusion approaches for bacterial classification based on spectral and sensor responses.Thirdly,the prediction and quantification of bacterial foodborne pathogen loads in pork samples from HSI and E-nose data were investigated.Further studies on developing an improved and efficient reduced spectrum model for quantitative tracking of foodborne pathogens was explored.Detection of BFP contamination of fresh pork with Escherichia coli O157:H7 and Staphylococcus respectively was implemented by employing HSI spectra information and partial least squares regression algorithm(PLSR).Variable selection algorithms applied for spectrum reduction included competitive adaptive reweighted sampling,genetic algorithm(GA),variable combination population analysis,ant colony optimization and iteratively retains informative variables algorithm.Variable combination population analysis with genetic algorithm hybrid strategy showed the highest residual predictive deviation(RPD)being 13.5910&16.8032,and best model performance of R_P~2=0.9977,0.9960;RMSEP=0.1532,0.1225;for E.coli and S.aureus respectively.Distribution maps provided a more insightful and detailed evaluation of the bacterial contamination at each pixel.Similarly,E-nose was applied for the identification,discrimination and quantification of S.Typhimurium contamination in pork samples.Principal component analysis(PCA)was successfully applied for discrimination of inoculated samples at different contaminant levels.After Optimization of the modelling parameters c and g for supper vector machine regression(SVMR)with PSO,GA and grid search algorithm,the.GA-SVMR model gave the best prediction accuracy for prediction set a studies and yielded an R_P~2=0.989;RMSEP=0.137;RPD=14.93.Lastly,the overall effectiveness of applying E-nose and HSI for in-situ discrimination and quantification of BFPs during the application of ultrasound technology for pathogen inactivation was evaluated.The main objective was to investigate and monitor the effects of ultrasound pretreatment for BFP inactivation in inoculated pork by HSI and E-nose.E-nose and HSI were employed as a process analytical technology,for nondestructive monitoring of ultrasound inactivation of S.Typhimurium,and E.coli.The contaminated pork was treated for 10,20 and 30 minutes.LDA model showed classification accuracies of 99.26%,and 99.63%,for S.Typhimurium,and E.coli using E-nose,and 99.70%,99.43%for S.Typhimurium and E.coli using HSI,respectively.PLSR quantitative models based on E-nose dataset showed Rp~2=0.9375,9.7240;RMSEP=0.2107,0.2057;RPD=9.7240,9.9604 for S.Typhimurium and E.coli respectively.Similarly,the model based on HSI dataset showed Rp~2=0.9687,0.9687;RMSEP=0.1985,0.2014;RPD=10.3217,10.1731 for S.Typhimurium and E.coli respectively.Overall,this study validates and pushes the frontiers of E-nose and HSI application for BFP detection in agar,broth and most importantly,in food matrices.This study further demonstrates the feasibility of applying a feature-level data fusion strategy on E-nose,HSI,and HS-GC-IMS datasets to produce better inferences than a solitary technique.From the conclusions drawn,and published results,it is feasible that the new approaches and ideas presented in this study could be adopted for controlling pathogen contamination in pork production or similar foods to ensure their safety.