Dissertation (Ph.D.)--University of Michigan.

Dissertation (Ph.D.)--University of Michigan.

Objective: The objective of this systematic review was to determine the feasibility and accuracy of machine learning (ML) versus a traditional statistical model in predicting dental caries in children. Methods: This is a systematic review comparing the feasibility and accuracy of ML versus traditional statistical models in predicting dental caries in children. The eligibility criteria were peer-reviewed studies published in English between January 1, 2007-December 31, 2022, that reported using both traditional statistical models and ML algorithms for caries prediction in primary dentition. The years 2007-2022 was chosen as the range because according to NIH.gov, in 2007, machine learning first started analyzing dental radiographs and images to look for tooth decay. Articles were extracted using search strategies from PubMed, Google Scholar, Embase, and Cochrane Library databases. Articles were screened using PRISMA guidelines, following a review for quality assessment using the JBI Critical Appraisal Checklists. Additionally, a meta-analysis was conducted comparing the studies that used traditional statistical models versus ML models using their pooled area under the curve (AUC) estimates. Only 3 of the 5 studies from each of the model types were analyzed based on the random effect estimate due to a limited number of studies. The meta-analysis was conducted using Med Calc software. Results: ML-based models that were most successful in predicting dental caries in children were Multilayer Perceptron (MLP) and Random Forest (RF). These algorithms outperformed the traditional statistical model of Logistic Regression (LR) as confirmed in the meta-analysis. However, some LR models outperformed certain ML models such as STVM and SVM and did not have much difference in predictive performance scores compared to ML algorithms such as XGboost. Low-income, frequency of dental visits, and toothache were the most significant factors that predicted caries among many of the studies. Low fluoride exposure, consumption of sugary food/drinks, and tooth brushing frequency were additional significant factors contributing to caries prediction. Our meta analysis showed that the pooled AUC scores of ML and statistical models were 0.808 and 0.776 respectively. Heterogeneity assessment of the 3 studies that used traditional statistical models meta analysis showed no significant heterogeneity while the 3 articles that used ML models showed significant heterogeneity (Higgin’s I2 test = 28% and 91%, respectively). A forest plot showcased the pooled AUC scores and a funnel plot showcased publication bias for each model. The test for publication bias showed that both statistical and ML algorithms had low bias. However, this could be inaccurate due to the limited number of studies. Conclusion: Machine learning is a highly plausible and successful method for caries prediction. Specifically, MLP and RF exceeded other ML algorithms and LR, in predictive performance. However, LR still outperformed or performed closely to some ML algorithms. Therefore, the best performing algorithms, MLP and RF, could be recommended as more robust and accurate analytical tools for caries predictions compared to LR, but LR also has predictive potential. / Oral Biology