Aplicação de ATR-FTIR em lágrimas humanas para discriminação em retinopatia diabética

• Autor(a) principal: Ono Júnior, Tadashi
• Data de Publicação: 2024
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFU
• Texto Completo: https://repositorio.ufu.br/handle/123456789/42059; http://doi.org/10.14393/ufu.di.2024.5056
• topic: Espectroscopia; ATR-FTIR; Lágrimas; Retinopatia diabética; Aprendizado de máquina; Spectroscopy; Tears; Diabetic retinopathy; Machine learning; CNPQ::CIENCIAS DA SAUDE; Ciências médicas; ODS::ODS 3. Saúde e bem-estar - Assegurar uma vida saudável e promover o bem-estar para todos, em todas as idades.

Introduction: Diabetes is currently one of the biggest public health challenges worldwide. Prolonged diabetes leads to various diabetic microvascular complications, such as retinopathy, nephropathy and neuropathy. It is likely that multiple factors are involved in predisposing diabetic individuals to these complications. Early detection or diagnosis is essential for developing strategies to reduce the risk factors and costs generated by diabetic complications. Objective: In this study, we applied ATR-FTIR spectroscopy combined with machine learning analysis to tear samples from individuals with diabetic retinopathy in order to differentiate them from diabetics without retinopathy and people without diabetes. The aim of this study is to test tears using a simple, fast, inexpensive and non-invasive method, which can be applied in a remote care setting, that is able to identify people with diabetic retinopathy. Materials and methods: We evaluated a total of 59 individuals, 24 in the non-diabetic group (ND), 17 in the diabetic group (D) and 18 in the diabetic retinopathy group (DR). Tears were collected using the microcapillary tube technique and stored in laboratory tubes under -80°C freezing. After thawing, 2 microliters were applied to an FTIR apparatus coupled with ATR and dried for 8 minutes. Spectral analysis of each sample was then carried out. Pre-processing and machine learning tools were applied in the study on the tear spectra generated by the ND, D and DR ATR-FTIR. Results: The best algorithm for discriminating between the ND and DR groups was AdaBoost, with an accuracy of 0.73, a sensitivity of 0.70 and a specificity of 0.76. Using SHAP (Shapley Additive Explanations), we identified that the wavenumbers 1151 cm- 1, 1425 cm-1, 993 cm-1, 1755 cm-1, 1287 cm-1, 1725 cm-1, 2970 cm-1, 2989 cm-1, 1129 cm- 1 and 1088 cm-1 were the most important in differentiating between the two groups mentioned above. The main corresponding biomolecular components are carbohydrates and lipids. With regard to group D compared to group DR, the best test was Artificial Neural Networks, which showed an accuracy of 0.73, a sensitivity of 0.77 and a specificity of 0.70. The most important wavenumbers in this method were 1256 cm-1, 1647 cm-1, 2899 cm-1, 1414 cm-1, 2879 cm-1, 1075 cm-1, 1179 cm-1, 929 cm-1, 916 cm-1 and 903 cm-1. The three most important biomolecular components are nucleic acids, proteins and lipids. Conclusion: The results obtained through prediction models, using ATR-FTIR spectroscopy applied to human tears associated with machine learning algorithms, point to the possible discrimination of subjects with diabetic retinopathy from those with diabetes and non-diabetics.