Jorjani Biomedicine Journal

Background and objectives: Application of statistical machine learning methods such as ensemble based approaches in survival analysis has been received considerable interest over the past decades in time-to-event data sets. One of these practical methods is survival forests which have been developed in a variety of contexts due to their high precision, non-parametric and non-linear nature. This article aims to evaluate the performance of survival forests by comparing them with Cox-proportional hazards (CPH) model in studying first birth interval (FBI).
Methods: A cross sectional study in 2017 was conducted by the stratified random sampling and a structured questionnaire to gather the information of 610, 15-49-year-old married women in Tehran. Considering some influential covariates on FBI, random survival forest (RSF) and conditional inference forest (CIF) were constructed by bootstrap sampling method (1000 trees) using R-language packages. Then, the best model is used to identify important predictors of FBI by variable importance (VIMP) and minimal depth measures.
Results: According to prediction accuracy results by out-of-bag (OOB) C-index and integrated Brier score (IBS), RSF outperforms CPH and CIF in analyzing FBI (C-index of 0.754 for RSF vs 0.688 for CIF and 0.524 for CPH and IBS of 0.076 for RSF vs 0.086 for CIF and 0.107 for CPH). Woman’s age was the most important predictor on FBI.
Conclusions: Applying suitable method in analyzing FBI assures the results which be used for making policies to overcome decrement in total fertility rate.

Background and Objectives: Currently, diabetes is one of the leading causes of death in the world. According to several factors diagnosis of this disease is complex and prone to human error. This study aimed to analyze the risk of having diabetes based on laboratory information, life style and, family history with the help of machine learning algorithms. When the model is trained properly, people can examine their risk of having diabetes.
Material and Methods: To classify patients, by using Python, eight different machine learning algorithms (Logistic Regression, Nearest Neighbor, Decision Tree, Random Forest, Support Vector Machine, Naive Bayesian, Neural Network and Gradient Boosting) were analysed. were evaluated by accuracy, sensitivity, specificity and ROC curve parameters.
Results: The model based on the gradient boosting algorithm showed the best performance with a prediction accuracy of %95.50.
Conclusion: In the future, this model can be used for diagnosis diabete. The basis of this study is to do more research and develop models such as other learning machine algorithms.