Prediction of amputation risk of patients with diabetic foot by artificial inteligence techniques

Diabetic foot ulcers due to diabetic foot syndrome are characterized by high morbidity and mortality and are among the most important complications of diabetes mellitus, which can have vital consequences such as amputation for patients.The main purpose of this doctoral thesis is to predict the amputation risk of diabetic foot patients using artificial intelligence-machine learning classification algorithms. The dataset used in the doctoral thesis includes 407 patients who were treated with the diagnosis of diabetic foot infection in Istanbul University Faculty of Medicine, Department of Underwater and Hyperbaric Medicine between January 2009 and September 2019 and their follow-up processes were terminated. In this research, diabetic foot patients were evaluated retrospectively and it is aimed to determine with artificial intelligence techniques the relationship between the demographic and clinical features of the patients at the time of admission, and the benefits of the treatment and amputation results obtained through follow-up. In the research, the process of Knowledge Discovery in Databases created by Fayyad, Shapiro and Smyth (1996) has been followed. The most important features related to the amputation risk of diabetic foot patients in the data set were determined through Principal Component Analysis. In this doctoral thesis, various prediction/classification models were created to predict the risk of both general – "overall" and major and minor amputation – "multi class" risk of diabetic foot patients. In the creating of these models, k-Nearest Neighbor algorithm, Naive Bayes Classifier, Classification and Regression Trees Algorithm (CART, Random Forest algorithm, Artificial Neural Networks, Extreme Learning Machines, Support xxx Vector Machines, Logistic Regression and XGBoost Algorithms were used. Genetic Algorithm and Particle Swarm Optimization algorithms were also used for hyperparameter optimization for specific algorithms. As model performance evaluation methods; cross validation, hold-out method, bootstrap sampling and three way split methods were used. The application of appropriate machine learning classification algorithms on the dataset and sub-dimension datasets was repeated both before and after the feature selection process. As a result, the models that show the best performance in the sub-dimension data set consisting of categorical and numerical values, which were created by feature selection; the "Random Forest Algorithm" (accuracy: 0,78; sensitivity: 0,67; specificity: 0,87 ; precision: 0,80 ; negative predictive value: 0,77; F-Measure:0,73) which is an based on ensemble learning methodology and "Logistic Regression" algorithms (accuracy: 0,90; sensitivity: 0,94; specificity: 0,86; precision: 0,84; negative predictive value: 0,95; F-Measure: 0,89), performed better than other algorithms in the defined experimental conditions. Therefore, a decision support system has been developed to help physicians and healthcare professionals dealing with diabetic foot treatment by integrating these two prediction models, which show the highest performance in the relevant experimental conditions to the R-Shiny user interface.