Deep Learning Improves Prediction of Periodontal Therapy Effectiveness in Chinese Patients.

The objective: This study aims to propose a new model to predict the specific treatment effectiveness at site level by analyzing massive amounts of periodontal clinical data with deep learning methods. The background data discussing the present status of the field: In light of the low accuracy of current tools, the proposed models cannot fully meet the needs of clinical effectiveness prediction and cannot be applied to on site level prognosis development and formulation of specific treatment plan. Materials and Methods: Periodontal examination data of 9273 Chinese patients were extracted and used to propose a Sequence-to-Sequence model after performing data management and reconstruction. The model was optimized by introducing the Attention mechanism. Results: In the test set, the model obtained an average site-level probing depth (PD) accuracy (defined as the proportion of sites with <1 mm deviation of the predicted result from the true value) of 92.4% and high sensitivity (98.6%) for the pocket closure variable. For sites with baseline PD <5 mm, the model achieved a prediction accuracy of 94.6%, while it decreased to 79.9% at sites with PD >= 5 mm. In contrast, for teeth with initial mean PD >= 5 mm, the prediction accuracy significantly differed between molars and non-molars. Conclusion: Our model is the first to predict the site-level effectiveness with high accuracy and sensitivity. Future prediction models should incorporate deep learning for improved clinical prediction.