Oana Vîrgolici

Background/Objectives: Obesity is a major risk factor for diabetes mellitus, a metabolic disease characterized by elevated fasting blood glucose and glycosylated hemoglobin levels. Predicting the percentage and absolute variations in key medical parameters based on weight changes can help patients stay motivated to lose weight and assist doctors in making informed lifestyle and treatment recommendations. This study aims to assess the extent to which weight variation influences the absolute and percentage changes in various clinical parameters. Methods: The dataset includes medical records from patients in Bucharest hospitals, collected between 2012 and 2016. Several machine learning models, namely linear regression, polynomial regression, Gradient Boosting, and Extreme Gradient Boosting, were employed to predict changes in medical parameters as a function of body weight variation. Model performance was evaluated using Mean Squared Error, Mean Absolute Error, and R² score. Results: Almost all models demonstrated promising predictive performance. Quantitative predictions were made for each parameter, highlighting the relationship between weight loss and improvements in clinical indicators. Conclusions: Weight loss led to significant improvements in dysglycemia, dyslipidemia, inflammation, uric acid levels, liver enzymes, thyroid hormones, and blood pressure, with reductions ranging from 5% to 30%, depending on the parameter.

Background: Automated data processing is creating and implementing technology that automatically processes data. This computer tool is recommended for doctors because it supports their everyday work, assists in medical diagnosis, and enhances patient care. The aim of this paper is to propose an informatic tool that can extract the values of some parameters of interest from blood test sheets in order to get medical alerts and to monitor the chronic disease. Methods: An application, written in Python, was developed Results: The computer tool extracted automatically the values of glucose, triglycerides, HDL-cholesterol, total cholesterol, and LDL-cholesterol from medical sheets (text-based file or graphic file, respectively), and saved them in a database, accessed and represented in a graphic form the most recent values of these parameters; alerts according to metabolic syndrome criteria and Framingham risk score were generated. Conclusions: This tool contributes to the management of the medical process, saving precious time and helping the doctor in detecting current or future health problems.

Type 2 diabetes is a metabolic disease that causes abnormal high levels of glucose in the blood. The pancreas is healthy, but the body doesn’t respond properly to its own insulin. The principal culprit is obesity, too much high fat tissue. So, measuring the body mass index or the waist circumference is a step to estimate the risk for this disease. Many people have no symptoms and the disease develops silently, causing serious problems with eyes, feet, heart and nerves. The prediction of diabetes is a very topical problem. In addition to medical guides, more and more machine learning models appear, trained on different databases. The purpose of these models is to predict diabetes, based on different parameters, not all of them coming from medical analyses. In the paper we present four diabetes prediction models, respectively based on the decision tree, support vector machine, logistic regression and k-nearest neighbors’ algorithms. All models are trained and tested on a database with approximately 65,000 records (divided into 70% for training and 30% for testing), which contains two blood markers (haemoglobin A1c and glucose), an anthropometric parameter (body mass index), age, gender and three categorical parameters (smoking status, hypertension, heart disease). We identify that Haemoglobin A1C and glucose are the most influential predictors. The models are evaluated in terms of accuracy score and confusion matrix and a ranking is presented at the end. The results obtained are very encouraging for all the presented models.

The need to increase employee performance and productivity has become vital in most companies nowadays, considering the number of changes that processes and people have faced during recent years in many organizations. This becomes even more important as it can sustain the growth of the company, as well as the competitiveness. This work will present multiple methods and comparisons between them for the process of building a machine learning algorithm to predict performance scores for employees in one organization; these methods include pre-processing the data, selecting the best variables, building the best algorithms for the available data, and tuning their hyperparameters. The current research aims to conclude on a collection of practices that will determine the best predictions for the given variables, so that human opinion can become less influential in employee appraisal, increasing objectivity and overall productivity.

In over 60% of patients with prediabetes, the evolution to diabetes can be stopped by changing lifestyle. The prevalence of diabetes mellitus was 11.6% and of prediabetes was 16.5% in the Romanian population aged 20-79 years. The deficiencies are not only at the health system level, but also in the application of prediabetes criteria existing in accredited guidelines. Although these guidelines imposed by the international diabetes federation are constantly updated, many doctors do not apply the recommended steps for diagnosis and treatment. We review, in the first part of the paper, several studies in predicting diabetes, which used different algorithms and techniques. In this second part of the paper, we propose a machine learning approach for prediabetes prediction, which uses kNN (k-Nearest Neighbors), DT (decision tree), SVM (Support Vectors Machines) and Logistic Regression (LR) algorithms. We used a dataset with 125 persons (men and women), with the following features: gender (S), serum glucose (G), serum triglycerides (TG), serum high-density lipoprotein cholesterol (HDL), waist circumference (WC) and systolic blood pressure (SBP). We used standardized medical criterion named Adult Treatment Panel III Guidelines (ATP III), which specifies that prediabetes diagnosis can be established if at least three of five parameters are outside the scale of their normal values. We obtained, for both algorithms, encouraging results in evaluating the models (in terms of confusion matrix, f1_score, accuracy_score).

In over 60% of patients with prediabetes, the evolution to diabetes can be stopped by changing lifestyle. Application of prediabetes criteria existing in accredited guidelines is very useful, representing an effective way to avoid prediabetes and diabetes. Although these guidelines imposed by the international diabetes federation are constantly updated, many doctors do not apply, mainly due to lack of time, the recommended steps for diagnosis and treatment. In this paper, a multi-layer perpeptron neural network model for prediabetes prediction is proposed, based on a dataset with 125 persons (men and women), with the following features: gender (S), serum glucose (G), serum triglycerides (TG), serum high-density lipoprotein cholesterol (HDL), waist circumference (WC) and systolic blood pressure (SBP). The output feature in the dataset (prediabetes or not) was based on a standardized medical criterion named Adult Treatment Panel III Guidelines (ATP III), which specifies that prediabetes diagnostic can be establish if at least three of five parameters are outside the scale of their normal values. Satisfactory results were obtained in evaluating the model.