| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.17 MB | Adobe PDF |
Advisor(s)
Abstract(s)
A diabetes é uma das principais causas de mortalidade e redução da expectativa de vida [1]. Os biossensores são uma tecnologia promissora para o diagnóstico de diabetes, oferecendo uma deteção rápida e precisa dos níveis de glucose no sangue. Nesta dissertação, foram desenvolvidos biossensores de glucose (Glu) de primeira geração utilizando elétrodos descartáveis de platina impressos por serigrafia (Pt-SPEs, do inglês,
Platinum Screen-Printed Electrodes). Inicialmente, os Pt-SPEs foram modificados com biografeno (BGr), um nanomaterial biocompatível, aplicado na superfície do elétrodo de trabalho. Em seguida, a enzima glucose oxidase (GOx) foi imobilizada na superfície do elétrodo modificado. Duas metodologias distintas foram estudadas: a adsorção química e o encapsulamento da enzima por eletropolimerização da dopamina na superfície do elétrodo. A influência do número de camadas de BGr e de GOx no desempenho analítico do biossensor foi avaliada. Para cada um dos métodos, foi realizada a caracterização eletroquímica por
voltametria cíclica (CV, do inglês, Cyclic voltammetry) e a caracterização morfológica dos elétrodos por microscopia eletrónica de varrimento (SEM, do inglês, Scanning electronic microscopy) e microscopia eletrónica de transmissão (TEM, do inglês, Transmission electron microscopy). O desempenho dos biossensores foi avaliado através da construção das respetivas retas de calibração em tampão fosfato salino (PBS, do inglês: Phosphate buffered saline) e soro Cormay (SC), além de testes de seletividade competitiva. Os dados foram obtidos pela técnica de cronoamperometria (CA) em dois tempos distintos, 60 e 120 segundos, respetivamente. Com os dados obtidos durante a construção e calibração dos biossensores em SC, foram testados vários algoritmos de Machine Learning (ML) para regressão não linear, visando obter um modelo capaz de prever com precisão a concentração de glucose e os parâmetros da reta de calibração de um biossensor de glucose. Para cada variável de saída estudada, foram treinados diversos algoritmos com todas as combinações possíveis de dados de entrada, selecionando-se o modelo que melhor descreveu os dados fora do conjunto de treino. O biossensor com melhor desempenho analítico, em termos de gama de linearidade, limite de deteção (LD) e seletividade, foi aquele que utilizou uma camada de BGr e imobilização de uma camada de enzima pela metodologia de encapsulamento, considerando os resultados para 60
segundos. A gama de linearidade da calibração realizada em SC diluído variou entre 0,75 mM e 40 mM, apresentando uma correlação de 0,988 e um LD de 0,078 mM. Entre os vários algoritmos testados, determinou-se que o algoritmo de Decision Tree é capaz de prever a ordenada da origem de uma reta de calibração com um coeficiente de determinação de 0,956. Para além disso, foram construídos modelos Random Forest capazes de prever o valor do declive e limite superior de uma reta de calibração com coeficientes de determinação de 0,812 e 0,900 e ainda provou ser promissor na previsão de concentração de glucose com um coeficiente de determinação de 0,738.
Diabetes is one of the main causes of mortality and reduced life expectancy [1]. Biosensors are a promising technology for diagnosing diabetes, offering rapid and accurate detection of blood glucose levels. In this dissertation, first-generation glucose (Glu) biosensors were developed using disposable screen-printed platinum electrodes (Pt-SPEs). Initially, Pt-SPEs were modified with biographene (BGr), a biocompatible nanomaterial, applied to the surface of the working electrode. Then, the enzyme glucose oxidase (GOx) was immobilized on the surface of the modified electrode. Two distinct methodologies were studied: chemical adsorption and enzyme encapsulation by electropolymerization of dopamine on the electrode surface. The influence of the number of BGr and GOx layers on the analytical performance of the biosensor was evaluated. For each of the methods, electrochemical characterization by cyclic voltammetry (CV) and morphological characterization of the electrodes by scanning electron microscopy (SEM) and transmission electron microscopy (SEM) were performed. The performance of the biosensors was evaluated by constructing the respective calibration curves in phosphate buffered saline (PBS) and Cormay serum (CS), in addition to competitive selectivity tests. The data were obtained using the chronoamperometry (CA) technique at two different times, 60 and 120 seconds, respectively. With the data obtained during the construction and calibration of the biosensors in SC, several Machine Learning (ML) algorithms for non-linear regression were tested, aiming to obtain a model capable of accurately predicting the glucose concentration and the parameters of the calibration curve for a glucose biossensor. For each output variable studied, several algorithms were trained with all possible combinations of input data, selecting the model that best described the data outside the training set. The biosensor with the best analytical performance, in terms of linearity range, detection limit (LD) and selectivity, was the one that used a layer of BGr and immobilization of a layer of enzyme using the encapsulation methodology, considering the results for 60 seconds. The linearity range of the calibration performed in diluted SC varied between 0.75 mM and 40 mM, presenting a correlation of 0.988 and a LD of 0.078 mM. Among the various algorithms tested, it was determined that the Decision Tree algorithm is capable of predicting the ordinate of the origin of a calibration curve with a coefficient of determination of 0.956. Furthermore, Random Forest models capable of predicting the slope value and upper limit of a calibration curve with coefficients of determination of 0.812 and 0,900 were constructed and further proved to be promising in predicting glucose concentration with a coefficient of determination of 0.738.
Diabetes is one of the main causes of mortality and reduced life expectancy [1]. Biosensors are a promising technology for diagnosing diabetes, offering rapid and accurate detection of blood glucose levels. In this dissertation, first-generation glucose (Glu) biosensors were developed using disposable screen-printed platinum electrodes (Pt-SPEs). Initially, Pt-SPEs were modified with biographene (BGr), a biocompatible nanomaterial, applied to the surface of the working electrode. Then, the enzyme glucose oxidase (GOx) was immobilized on the surface of the modified electrode. Two distinct methodologies were studied: chemical adsorption and enzyme encapsulation by electropolymerization of dopamine on the electrode surface. The influence of the number of BGr and GOx layers on the analytical performance of the biosensor was evaluated. For each of the methods, electrochemical characterization by cyclic voltammetry (CV) and morphological characterization of the electrodes by scanning electron microscopy (SEM) and transmission electron microscopy (SEM) were performed. The performance of the biosensors was evaluated by constructing the respective calibration curves in phosphate buffered saline (PBS) and Cormay serum (CS), in addition to competitive selectivity tests. The data were obtained using the chronoamperometry (CA) technique at two different times, 60 and 120 seconds, respectively. With the data obtained during the construction and calibration of the biosensors in SC, several Machine Learning (ML) algorithms for non-linear regression were tested, aiming to obtain a model capable of accurately predicting the glucose concentration and the parameters of the calibration curve for a glucose biossensor. For each output variable studied, several algorithms were trained with all possible combinations of input data, selecting the model that best described the data outside the training set. The biosensor with the best analytical performance, in terms of linearity range, detection limit (LD) and selectivity, was the one that used a layer of BGr and immobilization of a layer of enzyme using the encapsulation methodology, considering the results for 60 seconds. The linearity range of the calibration performed in diluted SC varied between 0.75 mM and 40 mM, presenting a correlation of 0.988 and a LD of 0.078 mM. Among the various algorithms tested, it was determined that the Decision Tree algorithm is capable of predicting the ordinate of the origin of a calibration curve with a coefficient of determination of 0.956. Furthermore, Random Forest models capable of predicting the slope value and upper limit of a calibration curve with coefficients of determination of 0.812 and 0,900 were constructed and further proved to be promising in predicting glucose concentration with a coefficient of determination of 0.738.
Description
Keywords
Diabetes Biosensor Glucose Oxidase Machine Learning Artificial Intelligence Nonlinear Regression Models Diabetes Biossensor Glucose oxidase Inteligência artificial Modelos de regressão não linear