MACHINE LEARNING MODELS FOR PREDICTING TREATMENT OUTCOMES IN PERIODONTAL THERAPY: LASER VS. CURETTAGE

Background: Periodontitis is a prevalent chronic inflammatory disease that can lead to attachment loss, bone
resorption, and tooth loss. Both traditional mechanical curettage and laser-assisted therapy are used in non-surgical
periodontal treatment, with studies indicating both modalities can effectively reduce periodontal pockets. Predictive
analytics, utilizing machine learning (ML), may help anticipate treatment outcomes and personalize periodontal
therapy. This study evaluates the accuracy of ML models in predicting treatment success in periodontal patients treated with either laser or curettage.
Materials and Methods: A retrospective dataset of 300 patients with periodontal disease was analyzed. Baseline
features included demographics (age, sex, smoking status, systemic conditions), initial clinical measurements (pocket
depth and clinical attachment level), presence of bleeding on probing or suppuration, and treatment type (laser vs.
curettage). The outcome Treatment_Success (successful vs. unsuccessful at 3-month follow-up) was the target
variable. Three ML models—Random Forest, Support Vector Machine (SVM), and Logistic Regression—were
trained and tested on stratified data. Performance was measured using accuracy, precision, recall, F1-score, and the
area under the ROC and precision-recall curves. Confusion matrices were created.
Results: All three models showed moderate predictive performance. The Random Forest achieved an accuracy of
0.70, a precision of 0.73, a recall of 0.94, and an F1-score of 0.82 in predicting treatment success. The SVM achieved
accuracy 0.73, precision 0.73, recall 1.00, and F1-score 0.85, while Logistic Regression had accuracy 0.68, precision
0.73, recall 0.88, and F1-score 0.80. The SVM’s high recall indicated a tendency to predict most cases as “success”
(sensitivity 100% but no specificity). ROC analysis revealed similar model discrimination (AUROC ~0.55–0.64),
and precision-recall curves reflected a class imbalance favoring successful outcomes. SHAP showed that baseline
disease severity (initial probing depth and attachment loss) and older age were associated with a higher risk of
failure. At the same time, treatment type (laser or curettage) and factors like smoking, systemic conditions, and
bleeding had smaller impacts.|
Conclusion: ML models can predict short-term periodontal treatment outcomes with fair accuracy. The Random Forest and Logistic models balanced sensitivity and specificity better than the SVM. Feature interpretability analysis suggests that initial pocket depth, attachment level, and age are key predictors of treatment success, aligning with known clinical risk factors. These findings underscore the potential of predictive analytics in periodontal therapy to identify patients at risk of poor outcomes and tailor interventions accordingly. Further validation on larger, prospective cohorts is needed.