Development and validation of an AI-driven tool to evaluate chewing function: a proof of concept.

BACKGROUND

Masticatory function is an important determinant of oral health and a contributing factor in the maintenance of general health. Currently, objective assessment of chewing function is a clinical challenge. Previously, several methods have been developed and proposed, but implementing these methods in clinics may not be feasible. Therefore, more efforts are needed for accurate assessment of chewing function and clinical use.

AIM: The study aimed to establish a proof of concept for development and validation of an automated tool for evaluating masticatory function.

METHODS

YOLOv8, a deep neural network, was fine-tuned and trained to detect and segment food fragments. The model's performance was assessed using bounding box recall metrics, segmentation metrics, confusion matrix, and sensitivity values. Additionally, a separate conversion test set evaluated the model's segmentation performance using physical units, demonstrated with Bland-Altman diagrams.

RESULTS

The YOLOv8-model achieved recall and sensitivity rates exceeding 90 %, effectively detecting and classifying food fragments. Out of 316 ground truth fragments, 301 were correctly classified, with 15 missed and 5 false positives. The Bland-Altman diagram indicated general agreement but suggested a systematic overestimation in measuring the size of post-masticated food fragments.

CONCLUSION

Artificial intelligence presents a reliable approach for automated analysis of masticatory performance. The developed application proves to be a valuable tool for future clinical assessment of masticatory function.

CLINICAL SIGNIFICANCE: The current study provides a proof of concept for development of an automated tool for clinical assessment of masticatory function.