Investigação de enviesamento de modelos de aprendizado demáquina para diagnóstico de doenças neurodegenerativasvia registros de marcha e voz

Abstract

This paper synthesizes the main findings of a study on biases in machine learning models for diagnosingneurodegenerative diseases through gait and voice analysis. We evaluated how oversampling techniques, such as gaitwindowing and the indiscriminate use of multiple voice samples per person, inflate performance metrics when treatingsamples from the same subject as independent. By comparing protocols that ignore or preserve these dependencies in twodatasets, we observed that disregarding them inflates the metrics, while preserving them provides more reliable assessments.The results emphasize the importance of proper segregation of samples to obtain diagnostic models.