Objectives: Silver diamine fluoride (SDF) is a nonsurgical therapy for the arrest and prevention of dental caries with demonstrated clinical efficacy. Approximately 20% of children receiving SDF fail to respond to treatment. The objective of this study was to develop a predictive model of treatment non-response using machine learning.Methods: An observational pilot study (N = 20) consisting of children with and without active decay and who did and did not respond to silver diamine fluoride provided salivary samples and plaque from infected and contralateral sites. 16S rRNA genes from samples were amplified and sequenced on an Illumina Miseq and analyzed using QIIME. The association between operational taxonomic units and treatment non-response was assessed using lasso regression and artificial neural networks.Results: Bivariate group comparisons of bacterial abundance indicate a number of genera were significantly different between non-responders and those who responded to SDF therapy. No differences were found between non-responders and caries-active subjects. Prevotella pallens and Veillonella denticariosi were retained in full lasso models and combined with clinical variables in a six-input multilayer perceptron.Discussion: The acidogenic and acid-tolerant nature of retained bacterial species may overcome the antimicrobial effects of SDF. Further research to validate the model in larger external samples is needed.