Predição de cárie dentária em crianças: uma abordagem de aprendizado de máquinas

• Ano de defesa: 2024
• Autor(a) principal: Reyes, Lilian Toledo
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/26339/00130000098rd
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Odontologia; UFSM; Programa de Pós-Graduação em Ciências Odontológicas; Centro de Ciências da Saúde
• Palavras-chave em Português: Aprendizado de máquinas; Cárie dentária; Crianças; Epidemiologia; Inteligência artificial; Prognóstico; Artificial intelligence; Children; Dental caries; Epidemiology; Machine learning; Prognosis; CNPQ::CIENCIAS DA SAUDE::ODONTOLOGIA
• Link de acesso: http://repositorio.ufsm.br/handle/1/31561

Few studies have addressed the influence of early childhood predictors on dental caries development throughout life. Artificial intelligence (AI) through machine learning (AM) has shown advantages in predicting health outcomes and can be useful for this purpose. Therefore, the general objective of this thesis was to assess the performance of ML algorithms in predicting dental caries in children. For this purpose, two articles were structured. The first systematically evaluated the success of ML algorithms in the diagnosis and prognostic prediction of dental caries. The second article sought to develop and validate caries prognostic models by implementing AM algorithms on data collected in early childhood. In the review, the main outcome was the performance of AM models (accuracy, sensitivity, specificity, AUC [area under the ROC curve (Receiver Operating Characteristic)]. Two independent reviewers selected the studies and performed the methodological quality assessment. In diagnostic studies (15), the AUC value ranged from 0.745 to 0.987. In the prognostic studies (5), the reported sensitivities were higher in those studies with low risk of bias (median [interquartile range, IQR] of 0.996 [0.971-1.000] vs. studies with uncertainty/high risk 0.189 [0–0.340]; p 0.025). We conclude that the use of MA is promising in the field although the overall applicability of the evidence was limited, due to concerns about risk of bias and use of data outside the actual clinical setting. The second article is a 10-year follow-up cohort. Data from 639 children aged 1 to 5 years, examined in 2010 who were reassessed in 2012 and 2020 were used (retention rate in the cohort: 73.1% and 66.9%, respectively). Demographic, socioeconomic, social capital, clinical and behavioral factors were collected through semi-structured questionnaires applied to the children's guardians at baseline. Clinical examinations were performed by previously trained and calibrated examiners. The main outcome was caries development during follow-up, assessed by the International Caries Detection and Evaluation System (ICDAS). Logistic regression (LR) models and others AI's own algorithms (decision tree, random forest and XGBoost) were applied, using RStudio software. At a 2-year follow-up, XGBoost achieved an AUC of 0.833, with caries severity being the strongest predictor. After 10 years XGBoost performed significantly better than LR (p<0.05) by DeLong's test. The SHAP algorithm, based on XGBoost, indicated factors such as previous caries experience, non-use of fluoride toothpaste, high frequency of sugar consumption, adverse socioeconomic conditions and weak social networks, as the most relevant factors in the prediction. Based on these studies, it is concluded that AM algorithms may perform better than traditional models in predicting caries. Although these results require external validation, the combination of AM and SHAP models shows potential for predicting caries development.