Evaluation of colour parameters at different preparation depths and prediction of tooth colour using machine learning

This study aimed to analyze colour changes occurring at different preparation depths and predict tooth colour using machine learning algorithms. Upper and lower anterior teeth ( n = 106), which were free of caries, restorations, and root canal treatment, were used in the study. The teeth were prepared to depths of 0.3, 0.5, 0.7, 1.0, and 1.5 mm from the vestibular surface. Colours were measured using a spectrophotometer (Vita Easy Shade V; Vita Zahnfabrik, Bad Säckingen, Germany). L * , a * , and b * values were recorded at each depth, and Δ L , Δ a , Δ b , and Δ E 00 values were calculated relative to the initial measurement. Depth-dependent L * , a * , and b * values were modeled using linear regression, polynomial regression, decision tree, random forest, and support vector regression models. The coefficient of determination ( R ²) values of all machine learning models were analyzed. Additionally, the colour changes across different depths were compared in the statistical analyses. The evaluation of the Δ L and Δ E 00 values revealed no statistically significant difference between depths of 0.3–0.5 mm and 1.0–1.5 mm ( p = 0.001). Except for the Δ L values at depths of 0.7 and 1.0 mm, statistically significant differences were found in Δ L and Δ E 00 values in the pairwise comparisons of other depths ( p = 0.001). The Δ L and Δ E 00 values at the dentin level were statistically higher than those at the enamel level. The Δ a levels at the dentin level were statistically lower than those at the enamel level ( p < 0.05). The Δ E 00 was >0.8 in most samples, even at a depth of 0.3 mm. All machine learning models yielded low R ² values. The effect of dentin became more pronounced with the increase in preparation depth; in particular, a significant increase was observed in the Δ L and Δ E 00 values. The Δ E 00 was >0.8 in most samples, even at a 0.3 mm preparation depth, demonstrating that even minimal reduction led to a noticeable colour change. However, the low R ² values across all machine learning models indicated that preparation depth alone was insufficient for predicting tooth colour.