Artificial neural networks in predicting impaired bone metabolism in diabetes mellitus

Abstract

Growing incidence of diabetes mellitus (DM), given significant socioeconomic consequences that low-trauma fractures entail, determines a need to improve diagnostic standards and minimize the risk of medical errors, which will reduce costs and contribute to better treatment outcomes in this category of patients.

Aim. To assess diagnostic capabilities of the method based on the use of an artificial neural network (ANN) for predicting changes in reparative osteogenesis in diabetes mellitus.

Materials and methods. A single-center, one-stage, cross-sectional study included 235 patients with type 1 and type 2 diabetes mellitus and 82 persons of the control group (the total of 317 patients). Further, the obtained data were processed using the MATLAB software to develop an ANN with a training (80%) and test (20%) set. The ANN model was trained by optimizing the relationship between a set of input data (a number of clinical and laboratory parameters: gender, age, body mass index, duration of diabetes mellitus, etc.) and a set of corresponding output data (variables reflecting the state of bone metabolism: bone mineral density, markers of bone remodeling).

Results. The ANN-based algorithm predicted estimated values of bone metabolism parameters in the examined individuals by generating output data using deep learning. Machine learning was repeated until the error was minimized for all variables. The accuracy of the validation test to predict changes in bone metabolism based on patient data was 92.86%.

Conclusion. The developed ANN-based method made it possible to design an auxiliary tool for stratification of patients with changes in bone metabolism in diabetes mellitus, which will help reduce healthcare costs, speed up the diagnosis due to fast data processing, and customize treatment for this category of patients.