Mandibular shape prediction model using machine learning techniques.

Abstract

To create a mandibular shape prediction model using machine learning techniques and geometric morphometrics. Six hundred twenty-nine radiographs were used to select the most appropriate craniomaxillary variables in different craniofacial pattern classifications using a support vector machine. To obtain the three-dimensional mandibular shape, a Procrustes fit was used on 55 tomograms, in which 17 three-dimensional landmarks were digitized. A partial least square regression was employed to find the best covariation between craniomaxillary angles and the symmetric components of mandibular shape. The model was applied to a new sample of six tomograms and evaluated by the mean absolute error. Each mandible predicted was assessed using the Hausdorff distance (HDu) and a color scale. The model was also exploratively applied to six new radiographs. Covariation was 88.66% with a significance of < 0.0001 explained by twelve craniomaxillary variables. Low differences between the original and predicted models were obtained, with a mean absolute error of 0.0143. The mean distance between meshes ranged from 0.0033 to 0.0059 HDu and each color scale demonstrated general similarity between the surfaces. This approach offered promising results in obtaining a mandibular prediction model that enhances shape properties in an economical way and is applicable to a Latin American population. Clinical proof of this method will require further studies with larger samples. This method offers a reliable, economic alternative to traditional mandibular prediction methods and is applicable to the Latin American population.