Research On Machine Learning Algorithms For Odor-sensing System Toward Assisted Diagnosis Of Wound Infection

The potential of odor-sensing approach for assisted diagnosis of wound infection is investigated in this dissertation.In the current wound infection odor sensing system,there are two key issues that need to be studied urgently.One is that the odor sensing technology is single,and there is a lack of research on the detection of wound infections based on multiple odor sensing technologies.The second is that the odor detection algorithm is relatively simple,and the detection accuracy is low,and there is a lack of special algorithms for the detection characteristics.Based on targeted design of wound infection odor sensing system and wound infection odor experiment,extensive algorithm researches on sensor array optimization for e-nose,biological recognition for Field Asymmetric Ion Mobility Spectrometry(FAIMS),information fusion for dual odor-sensing system are conducted in this dissertation.The major research achievements and innovations are as follows.(1)Development of odor-sensing system based on integration of e-nose and FAIMSSensing the odor of an infected wound is very difficult because of the extremely biological complexity and small difference.For this reason,the traditional e-nose technology could not accomplish this hard work separately due to its poor detection limit(ppm level)and susceptibility to interference.FAIMS is a newly developed mass spectrometry technique with advantages of miniaturized fabrication,fast response,low detection limit(ppb level),good specificity.Therefore,by combining e-nose and FAIMS technology,a dual odor-sensing system for the detection of wound infection is proposed in this thesis.In a detection cycle,the e-nose and FAIMS detection unit of the system are set to simultaneously sense the odor of wound sample by auto-control procedure,wherein the sensed odor data are stored in host computers.Based on equipment performance test of Chongqing Academy of Metrology and Quality Inspection and laboratory test to the wound odor of SD rats,it can be concluded that the odor-sensing system is reasonable and stable,and it is able to effectively sense the odor of three typical pathogens’ infected wounds and the uninfected wounds.(2)Design and realization of reliable wound infection odor experiment on SD ratsBy first introducing Full-thickness Wound Model for this research area to establish skin wound infection model,and using effective experimental measures to avoid microbial contamination,the targeted wound infection on SD rats was successfully achieved.Thereby,the odor collection experiment was accomplished with the acquirement of a reliable rat wound infection odor dataset.(3)Development of a novel sensor array optimization algorithm for e-nose based on feature group selectionAt present,it is common to employ the feature selection method from machine learning community to realize sensor array optimization for e-nose.However,it has ignored the characteristic of the working principle of e-nose,i.e.,the recognition ability of e-nose is established on the odor response pattern derived from a set of sensors which are of cross sensitivity and broad specificity.In consideration of this,a novel sensor array optimization algorithm named Feature Group Evolution(FGE)was proposed in this dissertation.Initializing from a feature rank according to the feature relevance,the FGE method stepwise selects features in groups with the mechanism of competition,crossover and mutation,and finally selects out the key feature set.Experiments demonstrated that the FGE method has the highlighted feature of algorithmic robustness.It has effectively achieved optimization in two bacterial odor datasets and six feature selection benchmark datasets,where the overall accuracy is obviously higher than the eight popular benchmarks.(4)Development of a novel FAIMS recognition algorithm based on spatial information integration and strategy of dynamic spectra analysisThe current FAIMS data analysis works generally adopt the strategy of stable spectra analysis,wherein each sample only scans and/or selects a few FAIMS spectra(usually one or two spectra)and the whole spectral area,or specific spectral area,or fixed spectral curves of the spectra is used for subsequent analysis.This strategy can avoid the interference signal in the data,but meanwhile,it may lost the effective information.Based on spatial information integration and strategy of dynamic spectra analysis,a novel FAIMS recognition algorithm named Local Warning integrated with Global Feature(LWGF)was proposed.The LWGF method respectively analyzes the data on global and local spatial scale,then establishes two basic models which are fused by feature fusion.Experiments show that,the LWGF method has excellent performance on clinical human wound infection dataset including six uninfected patients and twenty infected patients,where the best accuracy and AUC(Area Under Curve)have achieved96.15% and 0.98,which are higher than the best existing method of 22.3% and 0.30,respectively.(5)Design of a proper information fusion framework and development of model-reliability-estimation-based decision fusion algorithmConsidering the principle difference and data heterogeneity of the involved two detection units,feature selection was first conducted in data block of each unit,thereby separate recognition models were established and finally they were fused in decision level.Besides,a novel decision fusion algorithm named Clustering-of-Model-Expression-and-Proximity-based Model Reliability Estimation(CMEP-MRE),which is nonlinear,nonparametric and data driven,was proposed in this thesis.The CMEP-MRE method first measures the proximity of the test sample and clustered training samples in the model representation space of each basic estimator,then estimates the model reliability as the model weights to add the basis representations,and finally uses the additive representation to establish the recognition model.Algorithm experiments demonstrated that,the framework could effectively fuse the separate odor information of the two detection units,and it has achieved the accuracy of 94.63% and 88.05% on 24 and 48 hour data of the rat wound infection odor dataset,respectively.