Artificial Intelligence Models for Diagnosing Pulpitis in Adults Using a Modified Wolters Classification: A...

To evaluate the performance of an artificial intelligence (AI) system in diagnosing pulp conditions in adult patients, using a modified Wolters diagnostic classification.

A cross-sectional study was conducted. Data from 200 teeth of 200 patients were collected. A clinical evaluation was performed, and a dataset containing 21 diagnostic attributes was compiled. A modified Wolters classification was used for more objective and reproducible diagnostic criteria. AI models (Decision Tree (DT), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and XGBoost) were trained using a K-fold cross-validation approach. Hyperparameters were optimised, and model performance was assessed through precision, recall, F1-score and area under the receiver operating curves (AUCs). The final metrics were calculated over several iterations. To estimate model variability, bootstrap resampling was applied. Additionally, attribute importance was analysed using the permutation method.

The diagnoses were distributed in normal pulp (NP) (21.5%), initial pulpitis (IP) (13%), mild pulpitis (MIP) (21.5%), moderate pulpitis (MP) (25%) and severe pulpitis (SP) (19%). During validation, the models that achieved the highest performance were XGBoost (mean F1-score (0.85)) and SVM (mean F1-score (0.85)). Regarding ROC curve analysis, both models demonstrated AUC values exceeding 0.93 for all diagnostic classes. SVM reached perfect AUCs for SP and near-perfect values for NP (0.98), while XGBoost achieved a perfect AUC for NP and values above 0.90 for all other classes. Bootstrap analysis confirmed that XGBoost has the lowest variability and the best average values for precision (0.81), recall (0.81) and F1-score (0.80). On the other hand, pain to cold stimulus emerged as the most relevant attribute for diagnosis.

The AI-based system demonstrated significant potential for achieving accurate pulp diagnoses, contributing to more objective diagnostic decisions. Pain intensity in response to thermal stimuli and spontaneous pain was a key feature that allowed correct classification, highlighting its relevance in the diagnostic process.