Machine learning in the prognostic appraisal of Class III growth

Machine Learning (ML) is an emerging subfield of artificial intelligence with significant resources being applied to connect computer science, statistics, and medical problems. Currently, and even more so in the future, ML algorithms applied to the orthodontic specialty will offer sophisticated and automatic models able to process and synthesize data in ways orthodontists could never do themselves, and ultimately convert data into intelligent treatment actions. This work focuses on the usefulness of two ML methodologies, LASSO networks (Ln), and Boruta selection (Ba), to simplify information from different types of pathogenic processes leading to the worsening of skeletal Class III malocclusion. Cephalometric analyses of 144 Class III untreated subjects followed longitudinally during the growth process (4–19 years) were performed. After separating subjects into two subgroups of 116 with mild (M) and 28 with very serious (VS) unfavorable growth, cephalometric features were processed using Ba and Ln algorithms for feature selection and regularization. The selection procedure revealed the unexpected predictive importance of the combination of two often overlooked craniofacial variables, SN-PP and L1-MP angles. Ln regularization highlighted additional feature interactions between M and VS growing subjects. Thus, the appropriate removal of redundant cephalometric features from the dataset contributed to the detection of subjects affected by serious unfavorable craniofacial progression and revealed the unexpected prognostic value of some skeletal feature interactions.