Intelligent Detection For Food Impaction With Interpretable Fusion Network

To develop an intelligent acoustic detection method for food impaction with an interpretable time-frequency domain dual-feature fusion model and to verify the effectiveness of the method internally and externally. The 1200 volunteers were recruited from July 2024 to February 2025 (2024SLL043). The researcher used Oral-B dental floss (50 μm) through the interproximal contact area between the first and second molars three times, recording the passing sound with professional equipment. The sampling frequency was 16000 Hz and saved in WAV format. Finally, the dataset contained 2023 samples: impacted teeth (n=1002) and non-impacted teeth (n=1021), which were divided into the training and testing set (8:2). The study proposed an interpretable time-frequency dual-stream joint learning fusion network to detect food impaction, which mainly includes extracting acoustic time-frequency domain features, identifying feature distribution by the dual-stream residual network, fusing time-frequency domain feature difference, and layer-by-layer iterative optimization. The model was compared with the existing network model, junior and senior dentists, and the indicators included accuracy, sensitivity, and specificity. The internal validation showed that the proposed model outperformed existing network models. The external validation showed that our model was comparable to senior dentists and significantly higher than junior dentists (P<0.001). The study is the first to achieve accurate and intelligent detection of food impaction based on the network. This method might effectively prevent food impaction and improve medical quality.