Development of a Noninvasive Diagnostic Model Using Machine Learning for the Identification of Precocious Puberty in Girls

Purpose To develop a noninvasive artificial intelligence (AI) prediction model to identify central precocious puberty (CPP) in girls with suspected precocious sexual maturation. Methods Retrospective cohort study with girls evaluated for precocious pubertal development. The girls were grouped into two categories for machine learning model development: CPP and non-CPP. We selected the random forest algorithm for model development, in which the dataset was standardized before the cross-validation procedure. We also measured accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) of the four defined scenarios. The first scenario was composed of all selected variables. The second scenario contained only ultrasound data (pulsatility index, uterine volume, uterine fundus-to-cervix ratio, and ovarian volume). The third comprised only clinical variables (age, height, Tanner breast stage). The fourth scenario contained the three best variables (Tanner breast stage, pulsatility index, and ovarian volume). Results We included 76 girls (46 CPP and 30 non-CPP). The scenario that included three variables (uterine arteries mean pulsatility index, largest ovarian volume, and Tanner breast stage) was able to identify the CPP with a mean AUC of 0.970 ± 0.04, which was comparable to the scenario with all initially selected variables (AUC 0.950 ± 0.05), and the one with only ultrasound data (AUC 0.940 ± 0.05). These three models had a significantly higher AUC than the scenario that included only clinical variables (AUC 0.880 ± 0.08) (P < 0.05 for the comparisons). Conclusion We developed an accurate noninvasive AI prediction model to identify CPP in girls using only three variables obtained from physical examination and Doppler ultrasound.