Factors influencing the predictive performance of artificial intelligence for craniofacial growth

To evaluate factors influencing the prediction error of artificial intelligence (AI) that predict craniofacial growth and to identify an optimal AI training condition to improve the predictive performance of the AI model.

Original growth data were collected from the Mathews longitudinal serial growth study. From the original data consisting of 1257 datasets from 33 growing children of northern European descent, 60 data subsets were generated using random resampling procedures to include 12, 18, and 24 subjects, with data sizes of 100, 200, 300, 400, and 500 datasets. The resampling procedures were repeated four times. Each subset was used to train and create a total of 60 AI models. The prediction accuracy of these models was evaluated using growth prediction errors at the lower lip landmark, labrale inferius, as a benchmark indicator. The prediction errors of the 60 AI models were analyzed according to the number of subjects and data sizes.

Prediction error decreased as the data size increased. However, increasing the number of subjects within the growth data led to higher prediction errors. Notably, the increase in prediction error caused by adding more subjects was more substantial than the improvement achieved by increasing the data size.

The findings suggest that developing highly accurate AI-based craniofacial growth prediction models remains a significant challenge, even with extensive datasets.