Endocrine Abstracts

Introduction: Several investigations proved the presence of receptors for androgens, estrogens and progesterone in vocal folds. The most common endocrine disorder causing alterations in the aforementioned hormones concentrations is polycystic ovary syndrome (PCOS), which, according to several studies, can cause voice changes, especially deepening of its timbre. Over last few years, vocal changes accompanying many different disorders have been a subject of research with the use of machine learning (ML) - an algorithm-centered branch of artificial intelligence.

Aim: The aim of this work was the voice analysis in PCOS and its subgroup - PCOS with laboratory hyperandrogenism (PCOS-HA). The analysis of voice samples comprised evaluation of chosen acoustic features in terms of their ability to predict selected groups, as well as training the classifiers evaluating the probability of a given voice sample’s owner belonging to PCOS or PCOS-HA.

Materials and Methods: The first study group comprised 39 patients with PCOS, while control group included 56 healthy women. The PCOS-HA subgroup comprised 17 patients and 49 healthy women as a control group. All participants underwent anthropometric measurements, oral glucose tolerance test, hormonal profile assessment and transvaginal ovarian ultrasonography. All participants provided voice recordings, further submitted for analysis by ml.

Results: The acoustic analysis revealed the differences between study groups and their healthy counterparts in terms of several dozen of acoustic features. Two of them were associated with PCOS independently of age, free androgen index and fasting glucose concentration. The classifier evaluating the probability that the voice analysis of the recorded woman indicates her belongingness to the PCOS group was distinguished by the balanced accuracy equal to 74.4%, the sensitivity of 57.1%, the specificity of 91.7%, the precision of 80% and the area under the curve (AUC) of 0.810. Five acoustic features were associated with PCOS-HA independently of age, BMI and fasting glucose concentration. The classifier evaluating the probability that the recorded woman belongs to the PCOS-HA group was distinguished by the balanced accuracy equal to 85%, the sensitivity of 100%, the specificity of 70%, the precision of 57.1% and the AUC of 0.950.

Conclusions: PCOS and PCOS-HA influence vocal changes. The classifiers predicting the probability of PCOS, based on the voice analysis, do not fulfil the criteria of a useful screening test in case of no additional information about the recorded patient. The classifier predicting the probability of PCOS-HA, based on voice analysis, fulfils the criteria of a useful screening test.