Predicting Local Anesthetic Success in Symptomatic Irreversible Pulpitis: A Machine Learning Study

Overview

This study will develop and internally validate three machine learning models - logistic regression, random forest, and XGBoost - to predict local anesthetic (LA) success in patients undergoing endodontic treatment for symptomatic irreversible pulpitis (SIP). A large retrospective cohort of 4,390 consecutive adult patients treated at a single center (May 2014-October 2025) is being analyzed. The dataset was frozen in October 2025 for this analysis.

This study will be designed as a retrospective, single-center, cross-sectional analysis of prospectively recorded clinical data collected from patients presenting for endodontic treatment of teeth diagnosed with symptomatic irreversible pulpitis (SIP). Data will be extracted from clinical records maintained in the Department of Conservative Dentistry and Endodontics, Faculty of Dentistry, Jamia Millia Islamia, New Delhi, spanning consecutive eligible patients treated from May 2014 onwards. Patients will be eligible for inclusion if they are adults (≥18 years) presenting with a confirmed clinical diagnosis of SIP in a permanent tooth, based on spontaneous and lingering pain to thermal stimuli, positive pulp sensibility testing, and absence of radiographic periapical pathology. Patients will be excluded if they have an ASA physical status classification of III or higher, are pregnant, have a confirmed periapical lesion or pulp necrosis, have received a local anesthetic injection for the same tooth within the preceding 24 hours, or have a documented allergy to amide local anesthetics. The primary outcome will be binary local anesthetic (LA) success, defined as the achievement of adequate pulpal anesthesia using the initial anesthetic technique without any supplemental injection, operationalized as an intraoperative Heft-Parker VAS (HP-VAS) score of less than 54 mm with no supplemental anesthesia administered by the treating clinician at any point during the procedure. Candidate predictors - selected on the basis of biological plausibility and availability as structured fields in the clinical record - will include patient age (continuous), sex, current alcohol use, tooth type (six categories: maxillary incisors/canine, maxillary premolar, maxillary molar, mandibular anterior, mandibular premolar, and mandibular molar), pre-operative HP-VAS pain intensity, and preoperative medication use. The study will be conducted and reported in accordance with the TRIPOD statement, and the protocol will be reviewed and approved by the Institutional Research Review Committee prior to data extraction. All statistical analyses will be performed using Python 3.10, with model development conducted using scikit-learn (version 1.3) for logistic regression and random forest models, and the XGBoost library (version 2.0) for gradient boosting. The full analytic cohort will be randomly partitioned into a training set (70%) and a held-out test set (30%) using stratified random sampling to preserve the proportion of LA success and failure in both partitions; the test set will be reserved exclusively for final, unbiased model evaluation. Three classification models will be developed: an L2-regularized logistic regression, a random forest, and an XGBoost model, with hyperparameters for all three tuned using stratified five-fold cross-validation within the training set. Class weighting will be applied during training of all three models to prevent bias toward the majority outcome. Model interpretability will be assessed using SHapley Additive exPlanations (SHAP) values via the TreeSHAP algorithm for the tree-based models, and using standardized odds ratios for logistic regression. Discriminative performance will be evaluated on the held-out test set using the area under the receiver operating characteristic curve (AUC) with 1,000-bootstrap 95% confidence intervals; calibration will be assessed using calibration plots, calibration slope, calibration intercept, and Brier score; and clinical utility will be quantified using decision curve analysis comparing model-guided supplemental anesthesia against treat-all and treat-none default strategies across a clinically plausible range of threshold probabilities.

Conditions

Eligibility

Outcomes