Examination of biochemical markers in rats injected with diabetes with streptozotocin and prediction of diabetes disease with artificial intelligence (machine learning) algorithms

This thesis aims to develop artificial intelligence-based diagnostic systems based on biochemical blood parameters in experimentally induced diabetes rat models. In the study, the role of new biochemical parameters or parameter combinations in diagnosis was examined in addition to traditional diabetes markers and a unique classification approach was proposed in this direction. The study was conducted at Necmettin Erbakan University Experimental Medicine Research and Application Center using 22 Wistar albino rats; the diabetes model was created by subcutaneous injection with streptozotocin at a dose of 40 mg/kg administered at 3-day intervals. At the end of the 52nd day, a total of 22 biochemical parameters were measured, including serum glucose, HbA1c, liver enzymes, vitamin D, minerals and lipid profile. According to the study results, in the STZ-applied diabetes group, serum ALP, ALT, AST, bilirubin, iron, fructosamine, glucose, HDL, calcium, creatinine, magnesium, resistin, total cholesterol, triglyceride, urea and HbA1c levels were significantly higher (p<0.05) than in the control group; albumin, PTH, total protein and vitamin D levels were significantly lower (p<0.05). According to the ANOVA F-statistic, HbA1c, fructosamine and glucose were determined as the most effective variables in classification. T-test results revealed a low effect of phosphorus and zinc parameters. In the logistic regression analysis, ALP, iron, fructosamine, glucose and vitamin D stood out. Experimental studies have shown that machine learning algorithms can classify diabetes with 100% accuracy with the dataset used in this thesis. In addition to HbA1c, glucose and fructosamine, the association of ALP, iron and vitamin D parameters was also found to be important for the classification process. The fact that fructosamine and glucose are prominent in both regression analyses shows that these parameters are confirmed to be strong biomarkers and the findings obtained are consistent with the literature. The dataset was increased with the SMOTE method, and 100% accuracy was achieved with decision trees, Naive Bayes, k-NN and artificial neural network algorithms after feature selection. The results show that a multidimensional approach based not only on classical glucose parameters but also on parameters such as liver function tests, mineral and vitamin levels is effective in diabetes diagnosis. This study makes an original contribution to the literature in terms of being one of the first approaches integrated with machine learning algorithms, where biochemical markers are interpreted with multivariate analyses, and reveals the potential of artificial intelligence-supported diagnostic systems in experimental biomedical applications.