Machine learning methods in the differential diagnosis of difficult-to-classify types of diabetes mellitus

The course of difficult-to-classify types of diabetes mellitus (DM) (slowly developing immune-mediated DM of adults (LADA), monogenic forms of DM (MODY)) has common features with both type 1 DM (T1DM) and type 2 DM (T2DM), so often remain misdiagnosed. Errors in determining the type of diabetes lead to incorrect treatment tactics, which leads to poor glycemic control, the development of complications, a decrease in the patient's quality of life, and increased mortality. The key method for diagnosing MODY is sequencing of genes associated with this disease, and LADA is an immunological blood test in combination with the features of the clinical picture. However, the exact criteria for referring patients to these studies have not yet been determined. Performing these studies on all patients without exception with risk factors can lead to unjustified economic costs, and access to them is often difficult. In this regard, various automated algorithms have been developed based on statistical methods and machine learning (deep neural networks, “decision trees”, etc.) to identify patients for whom an in-depth examination is most justified. Among them are algorithms for the differential diagnosis of T1DM and T2DM, algorithms specializing in the diagnosis of only LADA or only MODY, only one algorithm is aimed at multiclass classification of patients with diabetes. One of the algorithms is widely used, aimed at diagnosing MODY in patients under the age of 35 years. However, existing algorithms have a number of disadvantages, such as: small sample size, exclusion of patients with MODY or older patients from the study, lack of verification of the diagnosis using appropriate studies, and the use of late complications of diabetes as parameters for diagnosis. Often the research team did not include practicing physicians. In addition, none of the algorithms are publicly available and have not been tested for patients in Russia. This manuscript presents an analysis of the main automated algorithms for the differential diagnosis of diabetes, developed in recent years.