Analysis of health records of diabetes patients treated in İstanbul with big data techniques

Objective: In our study, we aimed to investigate the prevalence of diabetes in Turkey and specifically in Istanbul, the main comorbidities in patients with diabetes throughout Turkey, and also to develop a model that predicts blood sugar control in those receiving treatment in Istanbul using machine learning methods. Method: The PHR system of the Ministry of Health of Turkey is e-Nabız. The data in the system is transferred to the Hadoop-based big data environment Cloudera (CDH, v.6.3.2) as de-identified. Queries were made on Cloudera with Apache Impala (v.3.2.0). We analyzed the data of cases diagnosed with diabetes in Istanbul in 2017 until 2020. Among those with at least one HbA1c value per year, the last two HbA1c values below 7 were grouped as controlled blood sugar and the others as uncontrolled blood sugar. We used Binary Logistic Regression(LR), Multilayer Fully Connected Neural Network(MLP), Random Forest(RF), and eXtreme Gradient Boost(XGB) to design models that can predict the group. Area under the curve (95% CI), sensitivity, selectivity, precision, F1 score, and accuracy were used to evaluate the performance of machine learning algorithms. Results: At the end of 2020 in Turkey, the prevalance of diabetic patients was 11.12% and 9.77% in Istanbul. In 2019, the number of age-weighted physician admissions per capita was 15.5 for diabetics, 9.5 for non-diabetics, and the number of prescriptions per capita was 7.9 for diabetics, and 4.5 for non-diabetics. Of the 77,724 diabetes records with sufficient data, 28,791 (37%) had blood sugar under control. Among the models, XGB was determined as the most successful method with 0.895 ROC AUC and 0.890-0.899 CI. LR, MLP, RF methods gave the results 0.889 (0.883-0.894); 0.887 (0.884- 0.892); 0.878 (0.873-0.883), respectively. Conclusion: It was seen that national electronic health records can provide important information on diabetes and many health problems. It is seen that the most successful method in estimating blood sugar control is XGB, and other methods also give satisfactory results. Keywords: diabetes mellitus, big data, machine learning, epidemiology.