Prediction of impaired consolidation of limb long-bone fractures using neural network analysis

Introduction Impaired reparative regeneration in patients with fractures is the most common complication; immunogenetic mechanisms play a leading role in its pathogenesis. Many researchers are engaged in the search for an "ideal" diagnostic marker. For this purpose, neural networks have been increasingly used, which allow not only to predict various pathological conditions but also to determine reliable options for prevention and treatment. The purpose of the study was to evaluate the effectiveness of predicting impaired consolidation of long-bone fractures of the extremities using the neural network data analysis. Material and methods We examined 108 young patients (WHO classification) with fractures of lower limb long bones. The clinical comparison group consisted of 62 patients without complications at the age of 34.5 [18; 44] years. The study group included 46 patients of similar age (36 [18; 44]) years and gender with delayed consolidation. The control group included 92 practically healthy individuals. Exclusion criteria from the study were any concomitant disease, other location and nature of injuries, alcoholism, as well as inaccurate reduction of bone fragments, and repeated operations. Patients who received antiresorption therapy and calcium supplements in the prehospital stage were also excluded. Laboratory (genetic) studies included determination of carriage of polymorphic molecules — TNFRSF11B-1181(G>C), IL6-174(C>G) , TGF β 1-25(Arg>Pro), EGFR-2073(A>T) and VDR(BsmI283G>A) . Amplification was carried out using primer sets Litekh-SNP (Russia). The risk of developing delayed consolidation was assessed using SPSS Statistics Version 25.0 (Neural Networks module). The predictive performance of the neural network was assessed using ROC analysis. Results For determining the importance of the independent variable, the following gradation was noted: TGF β 1-25(Arg>Pro) gene polymorphism — 100 %; gene polymorphism TNFRSF11B-1181(G>C) — 97.1 %, gene polymorphism VDR-BsmI283(G>A) — 34.7 %; IL6-174(C>G) gene polymorphism — 31.5 %; polymorphism of the EGFR-2073(A>T) gene — 15.3 %. The percentage of incorrect predictions was 8.3 %. Area under the curve of ROC analysis (AUC) = 0.91[0.85–0.98], p < 0.001. The specificity of the resulting model is 0.95 %, sensitivity is 0.87 %, accuracy is 91.7 %. Conclusion The use of the neural network for predicting delayed consolidation of fractures using data on the carriage of certain gene polymorphisms has a sufficient degree of accuracy (91.7 %), which indicates that the introduction of the neural network analysis into practical medicine is promising.