Agenesis Patterns in Orofacial Clefting using Interpretable Machine Learning.

HOWE BJ; HECHT JT; MARAZITA ML; ZENG E; ORIOLI IM; WANG T; POLETTA FA; MORENO URIBE, LINA M.

Vancouver

Congreso; 97th IADR/AADR/CADR General Session & Exhibition; 2019

IADR/AADR/CADR

Objectives: Children with orofacial clefts (OFC) often show a wide range of dental anomalies, adding complexity to understanding the phenotypic spectrum of OFC. The prevalence of tooth agenesis is known to be increased in subjects with OFC compared to controls. The evidence is minimal, however, on whether agenesis occurs in unique patterns and how ethnicity, gender, cleft type, laterality, and the presence of other dental anomalies factor into those patterns. This study aims to identify patterns of tooth agenesis in the largest international cohort to-date of children with nonsyndromic OFC and controls. If identified, such patterns might offer insights as to specific mechanisms that account for tooth agenesis in OFC.Methods: 1,720 individuals from 10 international sites were included: 887 subjects with OFC and 833 controls. Tooth agenesis was identified from in-person dental exams or intra-oral photographs. Patterns were identified using a unique interpretable machine-learning algorithm using Bayesian-Rule-Set that learns rule sets capturing the main characteristics/patterns for case-probands and control-probands with agenesis. Each rule captures a subgroup of subjects with agenesis and reflects different patterns. The algorithm automatically determines the significance of each rule. True positive (TPR) and false positive (FPR) rates were calculated for each pattern.Results: For case-probands, patterns identified cover 34%-64% of subjects and returned FPRs below 7% with cleft lip and palate, male gender, and rotation being prevalent factors. For control-probands, patterns identified cover 13%-58% of subjects and returned FPRs below 8% with male gender, rotation and study site being prevalent factors. Male gender and rotation were common factors in case and control patterns, though no exact pattern matches occurred.Conclusions: Our novel analytical approach identified combinations of features that suggest patterns that are predictive and descriptive of agenesis in case-probands and control-probands. Overlap of patterns between cases and controls are suggestive of similar genetic pathways in subgroups of subjects.