A Multimodal Model for Caries Screening Using Intraoral Images and Questionnaires.

INTRODUCTION AND AIMS: To develop a multimodal deep learning model for dental caries screening in children by integrating intraoral photographs and questionnaire-based data, and to compare its performance against that of a unimodal model utilising images alone. METHODS: In total, 7671 images (3913 occlusal and 3758 smooth surfaces) were collected from children across different dentition stages. Oral health questionnaires were also conducted. The images were clinically labelled into three categories: caries-free, early caries, and moderate-to-severe caries. A multimodal model integrating intraoral images and questionnaire data was trained and evaluated against an image-only unimodal model. The performance metrics included accuracy, precision, recall, F1 score, and the area under the receiver operating characteristic curve (ROC-AUC). SHapley Additive exPlanations (SHAP) analysis was conducted to interpret feature importance. RESULTS: For the multimodal model, occlusal surface classification achieved an accuracy of 92.3% for the caries-free category, 89.2% for the early caries category, and 93.7% for the moderate-to-severe caries category, with ROC-AUCs of 0.967, 0.911, and 0.972, respectively. The performance of smooth surfaces was similarly strong (an accuracy of 92.1% for the caries-free category, 96.7% for the early caries category, and 93.4% for the moderate-to-severe caries category), with ROC-AUCs of 0.938, 0.908, and 0.963, respectively. Incorporating questionnaire data improved recall for the early caries category without compromising overall accuracy. SHAP analysis provided interpretable insights into model decision-making. CONCLUSION: Compared with the image-only unimodal model, the multimodal deep learning model increased early caries recall while preserving overall diagnostic performance. CLINICAL RELEVANCE: This study describes a multimodal caries intelligent screening model, demonstrating its potential as a population-level screening tool. However, further large-scale, multicentre validation is necessary to confirm its generalisability and effectiveness.