Predictive modeling for step II therapy response in periodontitis - model development and validation.

Area of Science:

• Periodontology
• Machine Learning
• Biostatistics

Background:

• Periodontal therapy, including Steps I and II, is crucial for treating periodontal disease but exhibits variable success rates.
• Predicting treatment outcomes at a granular level is essential for optimizing patient care and managing expectations.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting changes in periodontal probing depth (PPD) following Step II periodontal therapy.
• To identify key clinical covariates influencing treatment outcomes at the patient, tooth, and site level.

Main Methods:

• Utilized patient-, tooth-, and site-specific clinical data to train machine learning models.
• Assessed model performance in predicting PPD changes, pocket closure, and response to therapy.
• Investigated the impact of model tuning and identified significant predictive features.

Main Results:

• Models accurately predicted the stability of healthy periodontal sites.
• Performance was suboptimal for diseased sites, though tuning improved accuracy.
• Key predictors identified include PPD, tooth-site characteristics, and tooth type.
• Pocket closure was predicted with fair accuracy, with baseline PPD being the most influential covariate.
• Models effectively predicted improvement in shallow pockets but underperformed for non-responding sites, with antibiotic treatment and tooth type being influential.

Conclusions:

• Machine learning models offer a foundation for site-specific outcome prediction in periodontal therapy, despite current limitations in predicting outcomes for all diseased sites.
• The models can stratify periodontal sites and estimate the probability of pocket improvement, aiding in patient communication.
• Further refinement of predictive models using clinical data holds potential for enhancing personalized periodontal treatment strategies.