AI-Assisted Evaluation of Dental Anxiety in Children: A Machine Learning Approach

Overview

This observational study evaluated whether children's dental anxiety could be identified from their speech using artificial intelligence and machine learning methods. Children aged 8-12 years attending a pediatric dentistry clinic answered a set of short, standardized questions before receiving dental treatment. Their speech was recorded, and their dental anxiety was assessed during the same session using three established measures: the Children's Fear Survey Schedule-Dental Subscale, the Modified Child Dental Anxiety Scale, and the Face Image Scale. Acoustic characteristics of the children's voices and linguistic characteristics of their spoken responses were analyzed together. Four machine learning algorithms were developed and evaluated to determine how accurately they could distinguish between children with lower and higher levels of dental anxiety. No treatment was assigned or modified as part of the study.

Dental anxiety can adversely affect a child's cooperation, treatment experience, and willingness to attend future dental appointments. Existing assessment methods primarily rely on self-report questionnaires, visual scales, and clinical observation. Artificial intelligence-based analysis of speech may provide an additional objective and non-invasive method for recognizing dental anxiety before treatment. This prospective, cross-sectional observational study included children aged 8-12 years attending the Department of Pediatric Dentistry at Marmara University Faculty of Dentistry between September 2025 and February 2026. Data were collected within the natural workflow of an active pediatric dentistry clinic. The study did not assign participants to any treatment or alter their planned dental care. Before dental treatment, each participant took part in a standardized 1-3-minute conversation conducted by the same researcher. The questions addressed everyday life, school, oral hygiene habits, previous dental experiences, and expectations regarding the planned dental visit. Speech was recorded using a standardized microphone position and recording protocol. Immediately after the speech recording and before treatment, dental anxiety was assessed using the Children's Fear Survey Schedule-Dental Subscale, the Modified Child Dental Anxiety Scale, and the Face Image Scale. Scores from each measure were categorized into lower and higher anxiety levels using established thresholds and served as reference labels for machine learning model development. All data were anonymized before analysis. The researchers' speech and prolonged silent segments were removed from the recordings, and the audio files were converted to a standardized format. Acoustic characteristics were extracted from the recordings using OpenSMILE. The spoken responses were transcribed and analyzed using natural language processing methods based on Whisper and a Turkish-language BERT model. Selected acoustic and linguistic features were combined into a multimodal dataset. Random Forest, XGBoost, LightGBM, and CatBoost algorithms were trained and evaluated using five-fold cross-validation. Model performance was assessed using accuracy, sensitivity, specificity, precision, F1 score, the area under the receiver operating characteristic curve, predictive values, and confusion matrices. Agreement between model predictions was also evaluated. The primary objective was to identify the machine learning approach that most accurately and sensitively distinguished children with lower and higher dental anxiety levels.