Assessment of Children's Dental Fear and Anxiety Using Deep Learning Techniques

Background About one-third of children face dental fear and anxiety, jeopardizing treatment outcomes. This study develops deep learning algorithms to automate drawing analysis, addressing the limitations of subjective and time-intensive manual assessments. Methods Two deep learning systems were developed. In the pilot study, drawings were manually labeled using the Child Drawing: Hospital score sheet. The first system used a two-step classification, focusing on human figures. A ResNet-18 model categorized drawings into high or low-average anxiety, while YOLO-v5 detected human figures, and YOLO-v8 classified those in low-average anxiety drawings. The second approach evaluated layout characteristics using pixel value calculations and scored color attributes with k-means clustering. A two-step deep learning process assessed human figure features: YOLO-v5 detected human figures and recognized faces, while YOLO-v8 classified faces as happy or unhappy. Final anxiety levels were determined by the combined feature scores. Accuracy and precision were reported for classifiers and object detectors, respectively. Results The first approach showed promising results, with the ResNet-18 classifier achieving a testing accuracy of 0.93 (95% CI: 0.878-0.982), the object detector achieving precision of 0.64 (95% CI: 0.531-0.749), and across five-fold cross-validation, the final classifier achieving a mean test accuracy of 0.66 (95% CI: 0.627, 0.6966). Confidence intervals were computed using the t-distribution with 4 degrees of freedom. Since the testing set was too small, a normal distribution assumption for the testing set was unrealistic. Therefore, a five-fold cross-validation was performed. The second approach achieved a final accuracy of 0.76, with individual features scoring a moderate accuracy of around 0.60. Conclusion This study demonstrates the potential of artificial intelligence to automate children’s dental fear and anxiety assessments through drawing analysis. Practical Implications: The developed algorithms can be used to screen children's dental anxiety, aiding clinicians in managing anxiety effectively, understanding patient concerns, and improving communication with children with linguistic or cognitive barriers.