Comparison of Supervised Machine Learning Models to Logistic Regression Model Using Tooth-Related Factors to Predict...

Area of Science:

• Periodontal disease research
• Biostatistics in dentistry
• Machine learning applications in healthcare

Background:

• Logistic regression (LR) is a common tool in dentistry for longitudinal study predictions.
• Nonsurgical periodontal treatment (NSPT) outcomes require accurate predictive models.

Purpose of the Study:

• To compare the predictive validity of supervised machine learning (ML) models against conventional logistic regression (LR).
• To assess the performance of ML models in predicting NSPT outcomes.

Main Methods:

• Trained five ML models (Random Forest, Decision Tree, Support Vector Classifier, K-Nearest Neighbors, Gaussian Naïve Bayes) using site-specific predictors.
• Evaluated models using data from 1108 sites, including bleeding on probing (BoP), probing pocket depth (PPD), and clinical attachment loss (CAL).

Main Results:

• Conventional LR model achieved 70.4% prediction accuracy, with PPD significantly associated with NSPT outcomes (OR=0.577, p=0.001).
• Gaussian Naïve Bayes (71.0%) and Support Vector Classifier (70.4%) showed comparable accuracy to LR.
• K-Nearest Neighbors (65.0%), Random Forest (62.0%), and Decision Tree (61.0%) had lower prediction accuracies.

Conclusions:

• Supervised ML models did not outperform the conventional LR model in predicting NSPT outcomes.
• Larger sample sizes and additional periodontitis predictors are required to enhance ML model accuracy for NSPT outcome prediction.