Abstract
BACKGROUND: Moving from the correlation between insulin-resistance and PCOS, metformin has been administered in some PCOS women improving ovulatory and metabolic functions and decreasing androgen levels. Inconsistency and unpredictability of response to metformin limit its extensive use. Aim of this study was to identify reliable predictors of response to metformin therapy for weight loss and reduction in plasma androgen levels using ANNs (artificial neural networks). METHODS: One hundred eight consecutive women with PCOS (ESHRE/ASRM 2003 Rotterdam criteria) treated with metformin 1500 mg/day, at inclusion and every 6 months underwent to a complete clinical, endocrine/metabolic assessment and ultrasonographic evaluation. Therapy outcomes were BMI reduction (≥1 kg/m(2)) in overweight/obese and free-androgen-index (FAI) decrease (≥1%) in hyperandrogenemic women. Semantic connectivity maps (SCMs) were obtained through Auto-CM, a fourth generation ANN, to compare patients' baseline clinical features to the treatment outcomes. Multivariate logistic regression analysis was used to assess the major predictor in drop-out patients and the associated risk. RESULTS: At 6 months 54 out of 103 (52,4%) obese patients showed BMI reduction and 45 out of 89 (50,6%) hyperandrogenemic women showed FAI decrease. The further response rates at 12 months were 30,6 and 47%, respectively. SCMs showed a clear polarization for both the outcomes with elevated accuracy. Treatment responsiveness resulted strictly related to oligo-amenorrhea and hyperandrogenemia at baseline. In addition, lower serum testosterone levels at baseline were found to be the major predictor of treatment discontinuation. CONCLUSIONS: In women with PCOS, menstrual pattern imbalance and ovarian androgens excess are the best predictors of metformin response. They may pave the way for a rethinking of the criteria for evaluating hyperandrogenism in order to better define the large population included in the diagnosis of PCOS. Baseline plasma testosterone level can serve as a sensitive marker to predict treatment compliance.