Aprendizado de máquina para o diagnóstico da hipersensibilidade dentinária utilizando espectroscopia no infravermelho por transformada de Fourier

Abstract

Dentin hypersensitivity poses a significant challenge to patients' well-being, requiring the use of non-invasive and comfortable diagnostic techniques. Vibrational methods, such as Fourier-transform infrared spectroscopy (FTIR), provide a swift and non-intrusive means of examining biochemical constituents within tissues. In this study, data from 40 samples of gingival crevicular fluid (comprising 20 samples with dentin hypersensitivity diagnoses and 20 control samples) underwent FTIR analysis, with preprocessing steps encompassing normalization via Amide I and application of Savitzky-Golay with derivatives. Several machine learning models, including KNN, Decision Tree, Naive Bayes, XGBoost, MLP, and BiLSTM, were developed and compared. XGBoost stood out with consistent results under Amide I normalization, demonstrating its robustness. Additionally, Savitzky-Golay preprocessing had a positive impact, especially on the BiLSTM model, highlighting the importance of preprocessing choice. These results support the feasibility of dentin hypersensitivity diagnostic models and the importance of proper data preparation to optimize machine learning algorithm performance.