Research On Key Techniques For Lateral Flow Immunoassay Based On Up-Conversion Nanoprobes

Point-of-care testing（POCT）,as an important role of the modern in vitro diagnostic industry,effectively reduces the cost of medical testing and makes the in vitro diagnostic scene begin to shift from hospital laboratories to primary care units.Among various commonly used instantaneous detection techniques,immunochromatographic sensing technology based on up-conversion nanoprobe,as the most important method for constructing up-conversion nanoprobe based lateral flow immunoassay（UCNP-LFIA）,has the advantages of low interference by endogenous fluorescence,good detection stability and low potential toxicity,which has been widely and deeply explored by researchers in recent years.However,the existing UCNP-LFIA still has the disadvantages of weak detection sensitivity and poor detection accuracy,which limits its popularization and application in the medical and testing markets.The signal intensity in the detection area of test strips and the ability of the sensor to recognize the up-conversion fluorescent feature signal are the main factors affecting the accuracy and sensitivity of UCNP-LFIA detection.To address the problems of low signal intensity in the detection area of existing UCNP-LFIA test strips and weak recognition of fluorescence signals by UCNP-LFIA,the dissertation researched a number of key techniques such as high-performance UCNP nanoprobe preparation,multi-dimensional recognition of UCNP-LFIA signals and multi-modal UCNP-LFIA signal integration computation.On the basis of these key techniques,a miniaturized UCNP-LFIA,which could be used for the sensitive detection of small molecule ketamine（KET）in human hair samples and the accurate quantitative detection of highly sensitive inflammatory marker C-reactive protein（CRP）in human fingertip blood samples was developed.The main contributions of the paper are as follows.Firstly,to address the problem of low signal intensity in the detection area of the test strips,the dissertation proposed a surface modification method for up-conversion nanoparticles（UCNP）based on mesoporous silica（m Si O₂）,which ensured the high luminescence intensity of the up-conversion nanoparticles and greatly improved their biocompatibility,thus effectively enhancing their immunosensing performance,eventually improving the fluorescence signal intensity in the detection area of the test strips.The preparation of high-performance up-conversion nanoprobes was accomplished by solvent thermal decomposition,organic templating,surface amino carboxylation modification,and cross-linking activation.Characterization experiments revealed that the developed UCNP@m Si O₂ probe had strong 542 nm visible green light emission under980 nm light excitation,uniform nanoparticle size,good dispersion,and the dispersion coefficient was lower than 0.1;the surface modification of the probe was appropriate,and the probe showed obvious functional group stretching vibration peaks in the infrared spectra,such as Si-O,Si-OH,C=O,etc.Compared with the UCNP probe with ordinary core-shell structure and the UCNP probe based on solid silica shell layer coating,the UCNP@m Si O₂ had higher immunosensing performance and biosignal conversion efficiency,which laid the foundation for the development of UCNP-LFIA.Secondly,the dissertation addressed the problem of low sensitivity of UCNP-LFIA detection,and researched the method of multi-dimensional recognition of up-converted fluorescent signals by inferential transfer learning,which effectively improved the discriminative ability of the sensor for weakly positive samples containing trace antigen.The paper selected the small molecule drug KET as the detection object,and prepared several low concentration KET standard antigen test strips for the establishment of the original dataset,and completed the training and validation of the transfer learning network.On this basis,the dissertation researched and constructed a lightweight transfer learning network for the problems of large size and difficult embedding of the existing network model,proposed an improved architecture of Efficient Net network based on large convolutional kernel,and developed a redetection algorithm to further improve the detection credibility.The validation results showed that the detection precision,recall and F1 score of the trained lightweight transfer learning network reached 99.05%,99.01%and 99.10%,respectively.The model size was only 43.8 MB,which was nearly 50%smaller than the original network architecture of Efficient Net,and solved the problem of high computational overhead and low detection efficiency of conventional UCNP-LFIA for low concentration samples.Thirdly,the dissertation addressed the problem of poor accuracy of UCNP-LFIA detection,and researched the method of multimodal computation of up-conversion fluorescence signals through model fusion Stacking ensemble learning,which effectively improved the accurate classification ability of the sensor for samples with similar concentration.The paper selected CRP,a highly sensitive inflammatory indicator in human body,as the detection object,and prepared 11 CRP standard antigen test strips for the original data set establishment.On the basis of completing the training of 10 network models,the dissertation proposed several multi-network ensemble learning models based on different voting mechanisms to address the problems of poor detection stability and high risk of overfitting of a single network,and verified and compared their detection performance.The experimental results showed that the seven-network ensemble learning models of Reg Net,Efficient Net,CSPNet,Shuffle Net v1,Shuffle Net v2,Res Net Xt and Res Net＿101 based on the weighted soft voting mechanism had the highest accuracy rate of 99.69%in the face of the similar samples detection,and the required computational resource consumption,detection time and storage space all met the practical usage requirements.In addition,the ensemble model had high detection accuracy and excellent anti-interference ability under the environment of multivariate noise,which laid a foundation for the application of UCNP-LFIA to real blood sample detection.Finally,based on the research results obtained at the level of probe preparation and signal processing algorithms,the dissertation constructed a miniaturized UCNP-LFIA for the highly sensitive screening of KET,a small-molecule drug in human hair samples,and the precise quantitative detection of CRP,a highly sensitive inflammatory indicator in human fingertip blood samples.The real sample validation results showed that the sensor had excellent intra-batch and inter-batch precision for KET and CRP detection,with coefficients of variation less than 10%,as well as good analytical specificity.In the comparison experiments with the existing gold standard detection methods,the developed sensor was able to screen out all the KET weakly positive samples with high sensitivity and the false negative rate was 0.When facing the negative samples with KET concentration near the threshold,the sensor caused 8 cases of misclassification,and the false positive rate was 8%.Meanwhile,the sensor showed high accuracy in the detection of 100 blood samples with similar CRP concentration,with the linear fit R² between the detected value and the true value being higher than 99%,and the overall false-positive rate being 0.02.Compared with the sensors based on ordinary fluorescent nanoprobes and similar up-conversion nanoprobes,the proposed UCNP-LFIA had a much higher sensitivity and accuracy,and required less detection time and lower computational resource expenditure,possessing broad prospects for application in the POCT market.