Treatment selection and anovulation predictors in patients with obesity and oligo-/amenorrhea assisted by neural network technology

Introduction . Approximately 20 % of reproductive age women are obese, and more than half of them experience menstrual irregularities, anovulation, and infertility. Aim: to determine predictors of ovulation restoration and to develop a model for individual treatment selection in patients with obesity and oligo-/amenorrhea based on neural network technology. Materials and Methods . The prospective randomized controlled study included 80 women – patients with obesity and oligo-/amenorrhea, divided into 2 groups, who received the following therapy for 6 months: 40 patients (group I) – a combination of myoinositol, D-chiroinositol, folic acid and manganese, the other 40 patients (group II) – metformin. After treatment, patients from both groups were divided into two clusters based on the "anovulation/ovulation" criterion. Anthropometric parameters were determined, pelvic organs ultrasound was performed, laboratory tests (indicators of carbohydrate and fat metabolism, amino acids and peptides, hormonal status, blood cytokine levels, inflammation markers, and blood micronutrient composition) were performed. To create a model for predicting ovulation restoration, a multilayer perceptron procedure was used. The diagnostic value of the prognostic model was determined using ROC analysis. Results . The average age of the patients was 27.9 ± 3.8 years, with the average body mass index (BMI) 33,4 [31.2; 34.0] kg/m2. The following parameters were identified as significant anovulation predictors using neural network analysis: waist-to-hip ratio (WHR), menstrual cycle duration, Homeostasis Model Assessment of Insulin Resistance (HOMA-IR), C-reactive protein, leptin, follicle-stimulating hormone, 25-hydroxycalciferol (vitamin D), and tumour necrosis factor alpha. Because these parameters reflect the metabolic profile in patients, interventions to restore ovulation should primarily be aimed at correcting it. The developed model for predicting the onset of ovulation has a sensitivity of 100 %, specificity of 80 %, accuracy of 93.8 %; the area under the ROC curve is 0.985 (p < 0.001), which allows to consider it sufficiently informative for specific drug selection. For practical purposes, an online calculator for individual drug selection has been developed (accuracy – 91.7 %). Conclusion. An integrated approach based on neural network analysis of study parameters available for wide clinical practice is promising for predicting the onset of ovulation while using a specific drug due to its high information content.