Analysis of data-driven diabetes subgroups using machine learning for prediction of glycemic control and complications

Abstract

Diabetes is a long-term illness that affects the metabolism and causes high blood sugar levels. This can result in severe harm to vital organs such as blood vessels, the heart, kidneys, eyes, and nerves over time. Type 2 diabetes is the most prevalent form, typically found in adults, and it develops when the pancreas is unable to produce enough insulin or when the body is unable to effectively use the insulin it produces, resulting in insulin resistance.

Machine learning principles have been applied to develop algorithms that support predictive models for the risk of developing certain diseases, including diabetes, which has become a global pandemic. The management of this disease following a uniform treatment algorithm is usually linked to evolved treatment failure and the development of diabetic complications, so it has become necessary to create algorithms that consider the clinical characterization of the patient, which would be possible due to the recent advancements in the knowledge of the genomic architecture of diabetes and its complications.

This research work aims to use machine learning techniques for classification to study the subgroups, generated through clustering techniques proposed by previous studies, to develop a continuation that allows the finding of patterns within subgroups for the generation of a model able to predict complications and the response to certain treatments of a patient given its clinical characterization, that is phenotypic data. This was done in collaboration with the Unidad de Investigación de Enfermedades Metabólicas (UIEM), which is part of the Instituto Nacional de Ciencias Medicas y Nutrición Salvador Zubirán (INCMNSZ), which provided data from cohorts of diabetic patients, as well as medical knowledge and feedback from health professionals.

In this study, five clusters or data-driven diabetes subgroups were identified in a Mexican cohort based on six clinical characteristics. It is shown that it is possible to obtain similar subgroups to those found in other studies by using METS-IR and METS-VF instead of HOMA-IR and HOMA2-β formulas. These subgroups can be associated with different risks of complications, and progression rates from treatments, which might have relevance in advancing toward precision medicine. However, after including these subgroups in the training of classification models to predict glycemic control and assess the presence of complications, the performance of the models remained similar, that is no noticeable improvement was presented.

This study showed that it is possible to build models that predict glycemic control after three months of initiating treatment with good performances, as well as create models that classify the presence of neuropathy. On the other hand, the generation of classification models for nephropathy and retinopathy did not yield acceptable performances.