Analysis of type 1 diabetes verbal autopsy data by machine learning techniques

Abstract

Big data is a term used for data sets with large, diverse and complex structures that are often quite difficult to analyze or visualize using traditional computing methods and approaches. Machine learning (ML) techniques are effective in analyzing these types of data and extracting information from them. Health care systems generate large amounts of data from record keeping and this supports a wide range of medical decisions like population health surveillance and disease management for the overall improvement of the quality of health care delivery. In areas where there are no health registration systems a method of verbal autopsy is relied on to give information of a likely cause of death. In this study type 1 diabetes (T1DM) verbal autopsy data from MRC/Wits Rural Public Health Transitions Research Unit (Agincourt) was used as a test case for applying modern machine learning classification techniques to ascertain the cause of death by type 1 diabetes. Machine learning techniques used for the classification task were artificial neural networks (ANNs) and random forests which are realized with a keras front end and tensor flow. Machine learning algorithms automatically learn to make accurate predictions based on past observations by learning patterns in the data for this study, they learned features present in patients with diabetes and were able to identify patients who could have died from the disease. This is the first study on type 1 diabetes verbal autopsy data by the two machine learning techniques in South Africa. Performance metrics like precision, recall, confusion matrix were used for these classifiers because the data was incredibly skewed and the results obtained show that the random forest classifier classified the deaths by diabetes better than the artificial neural network. In particular the roc-score compares favourably with the study that was done by two clinician specialists in the disease whose study was similar