Development of an Intelligent Model for Prediction of Periodontitis Stage and Grade

• Author / Creator

  Ameli, Nazila

• The published and unpublished data in this thesis address a crucial gap in the utilization of comprehensive automated methods for diagnosing periodontitis. This research focuses on the development of a clinical decision support system (CDSS) for analyzing text and image data to diagnose periodontitis, aiming to enhance accuracy and efficiency in periodontitis stage and grade classification. Periodontitis, a complex inflammatory disease affecting dental supporting tissues, poses significant risks to oral and systemic health, including tooth loss and associations with conditions like diabetes and coronary artery disease. Early detection is crucial to mitigate its impact and avoid invasive treatments. My thesis aims to predict periodontal disease stage and grade by integrating novel methods for measuring bone loss (BL) with natural language processing (NLP) analysis of patients' chart data.In Chapter 2, utilizing Bidirectional Encoder Representations from Transformers (BERT), an automated methodology was developed to extract critical information from textual clinical notes for the classification of periodontitis stages and grades. This process involved fine-tuning the BERT model on a dataset of 309 clinical notes, which encompassed a diverse range of unstructured and narrative texts containing patient histories, medical and dental histories, possible treatment plans, and diagnostic observations.Through meticulous pre-processing, including tokenization and normalization, the clinical notes were adequately prepared for the BERT model. The fine-tuned BERT model demonstrated a remarkable ability to accurately identify and classify periodontitis stages and grades based on the extracted information. The evaluation of the model's performanceindicated high accuracy, showcasing its potential to streamline the diagnostic process by providing precise classifications from unstructured clinical notes.In Chapter 3, advanced deep learning (DL) models were utilized to analyze periapical (PA) radiographs, automating the process of BL segmentation and quantifying BL percentage relative to root length for predicting the stage and grade of periodontal disease. This approach involved the integration of U-Net and YOLO-V9 models to address specific aspects of the diagnostic process.The U-Net model was trained to perform precise segmentation of BL areas within the radiographs. By leveraging its convolutional neural network (CNN) architecture, U-Net demonstrated a high level of accuracy in delineating the boundaries of BL regions. This automated segmentation provided a reliable foundation for further analysis. Concurrently, the YOLO-V9 model was utilized to detect and localize the coordinates of the apex of teeth within the radiographs. The YOLO-V9 model's performance indicated that its detection capabilities were comparable to those of the specialists. This ensured that the model could accurately identify critical anatomical landmarks necessary for subsequent measurements. By combining the outputs of these models, the maximum BL was measured as a percentage of root length. This quantitative assessment was then used to classify the stage and grade of periodontal disease. Through this chapter, it became evident that the U-Net model was successful in the correct segmentation of BL, and the YOLO-V9 model effectively detected the coordinates of the tooth apex. The integration of these models facilitated accurate classification of periodontal disease stages and grades, highlighting the potential of DL techniques in enhancing periodontal diagnostics and treatment planning.Finally, in Chapter 4, by integrating these approaches through a multimodal transformer model, a comprehensive automated diagnosis system is proposed that enhances diagnostic accuracy and enables timely interventions to prevent adverse outcomes and costly treatments. This innovative system combines the strengths of textual clinical note analysis and PA radiograph examination through a multimodal transformer model.By combining these modalities, the multimodal transformer model synthesizes textual and visual data, enabling an accurate assessment of each patient’s condition. The model focuses on patients with the highest amount of BL, as identified through radiographic analysis, and correlates this data with the contextual information from clinical notes to classify the stage and grade of periodontal disease more accurately.The findings demonstrated that this integrated approach significantly outperforms previous single-modality models. While BERT was highly effective for text processing and the U-Net and YOLO-V9 models excelled in image processing, the model achieved a higher classification accuracy by combining these methodologies. The multimodal transformer model's ability to analyze and integrate diverse data types leads to a more precise and comprehensive diagnosis, facilitating timely and effective interventions in periodontal disease management. This advanced system not only improves diagnostic outcomes but also has the potential to reduce the need for extensive treatments by enabling early and accurate identification of periodontal disease stages and grades. In conclusion, this work demonstrates the potential for developing CDSS tools to improve periodontal disease management and patient outcomes.

• Subjects / Keywords

  • Periodontitis
  • BERT
  • U-Net
  • YOLO
  • Machine learning
  • Multimodal

• Graduation date

  Fall 2024

• Type of Item

  Thesis

• Degree

  Doctor of Philosophy

• DOI

  https://doi.org/10.7939/r3-7fj4-mn54

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.