Diagnosis of Methylmalonic Acidemia using Machine Learning Methods

Methylmalonic acidemia (MMA) is an autosomal recessive metabolic disorder. Traditional diagnosis needs physicians' personal level of professional medical knowledge and clinical experience. In this paper, we employ machine learning methods to diagnose MMA based on patients' laboratory blood tests and laboratory urine tests, in order to make a timely diagnosis and reduce dependence on physicians' personal level of professional medical knowledge and clinical experience. By comparing different machine learning algorithms for diagnosing MMA, we obtain the following conclusions: (a) machine learning methods can perform well for diagnosing MMA (all established predictive models obtain high accuracies and AUC values which are greater than 0.85 over all data sets, and some of these results are even more than 0.98); (b) random forest algorithm performs best among the compared algorithms; and (c) diagnosis based on the data combining both urine tests and blood tests is better than diagnosis based on single test alone in general. The conclusions show that applying machine learning algorithms to the diagnosis of MMA can achieve good performance. Thus, it is credible to build machine learning models to give an initial diagnosis without professional medical knowledge.