A Predictive Model for Mortality of Patients with Thalassemia Using Logistic Regression Model and Genetic Algorithm

Due to the thalassemia severe complications, prediction of mortality or patients survival has a great importance in early treatment phases. This study purpose was to predict the mortality rate of patients with thalassemia major and thalassemia intermedia, by the use of the binary logistic regression algorithm and genetic algorithm combination.

This retrospective cohort study was conducted on 909 thalassemia patients by using a questionnaire during 2004-2014. The data of all patients referring to Imam Reza Hospital from 2004 to 2014 have been considered. This study predictive variable is considered to be death or survival of the patient. In this research, we embedded the missing data by the use of the proposed data mining model and MICE algorithm. Totally, 100 patients were excluded from this research, due to the missing or out-of-range data. Death was considered as dependent variable. Also, a predictive model was designed in order to
predict the patient mortality using MATLAB language.

Mean age of the thalassemia patients was 25.7±9.04 years old and at the end of the study death was reported in 185 subjects. Additionally, there were also 26 independent variables. Moreover, the missing variables mean for each patient was 1.8±0.81. The combined predictive model was able to predict the patient survival rate with 94.35% accuracy. In this research, it was found out that 26 independent variables, which were collected from 12 variables were patient mortality predictors. Also, missing data imputation is an important method for increasing the data mining algorithms efficiency.

According to this study results, the use of missing algorithm with the data analysis aid yielded more accurate results, in comparison with the MICE algorithm. Furthermore, 12 parameters affected the patient mortality prediction, which were extracted by the genetic algorithm. Accuracy of the predictive model for the patient death detection was favorable. Consequently, it is recommended to use this model in order to predict the patient mortality.