Rastreamento de ansiedade por meio da voz em gestantes utilizando machine learning

Abstract

This work addresses the issue of the difficulty in accessing early diagnosis of anxiety in pregnant women, which is often delayed due to barriers in accessing healthcare professionals. This limitation highlights the need for accessible and efficient alternatives to support the identification of anxious signs through non-invasive methods. A total of 110 audio samples and demographic data from the Grávida Digital project were used as input for the models, with the GAD-2 and PSS questionnaires employed as reference for labeling the data. The aim of this study is to track signs of anxiety in pregnant women through voice, using Machine Learning techniques. The methodology adopted was an exploratory research based on previously collected data. The process involved the analysis and processing of data, extraction of acoustic features, and audio processing, as well as evaluating different classification models to identify the most effective approach. Feature sets such as MFCC, OpenSMILE, and PyAudio (including MidTerm) were tested, applied to 8-second windows. For classification, five algorithms were evaluated: KNN, Logistic Regression, SVC, XGBoost, and Gradient Boosting, using metrics such as AUC, F1-score, and accuracy. The tools employed include Python, scikit-learn, Librosa, and OpenSMILE for data extraction and processing. The results indicated that the MFCC feature set was the most robust, and that the inclusion of demographic data can improve the model performance. The best performance was achieved with the Logistic Regression model, using PyAudio with GAD-2 scores as labels, reaching an AUC of 0.705 and F1-score of 0.51. However, the variation between classifiers suggests that ensemble techniques and parameter adjustments may improve the stability of the models. It is concluded that voice analysis has the potential to track anxious signs in pregnant women, which can be further enhanced with additional optimizations.