IADR Abstract Archives
Application and Validation of Machine Learning Algorithms In Periodontal Disease.
Objectives: To apply and validate machine learning algorithms to dental periodontal findings, diagnosis and treatment.
Methods: Machine learning-based dental imaging solutions are complex and can easily be misrepresented, so we have designed a set of core principles, optimal reporting practices, and experimental design that encourage standardization and interpretability. We utilized data elements for machine learning modeling building and refinement, from the electronic health record. We extracted standardized periodontal disease findings, diagnosis and treatment, attachment loss, gingivitis, and localized and generalized severe periodontitis. Clinician determined validated cases (n=5,180) were extracted with all EHR periodontal chart data, findings, diagnosis, procedure and digital radiographs for each case. 80% of the sampling data was used to apply the machine learning algorithms with interpretation of the radiographs for cases with good periodontal health as compared to patients with periodontitis (n=4,144). The remaining 20% of the extracted data was used to validate the system's ability to interpret radiographs without EHR elements (n=1,036).
Results: Of the 4,141 patient records, 96.6% were successfully utilized for machine learning algorithms. Data elements of 927,616 were successfully extracted and utilized. 14,494 of dental radiographs were utilized. The machine learning algorithm was able to consistently and accurately determine periodontal pocket depth with corresponding periodontal diagnosis and appropriateness of care.
Conclusions: By applying electronic health record derived periodontal findings, diagnosis and treatment in conjunction with dental radiographs, machine learning-based algorithms were able to accurately determine periodontal findings, diagnosis and appropriateness of care.
Methods: Machine learning-based dental imaging solutions are complex and can easily be misrepresented, so we have designed a set of core principles, optimal reporting practices, and experimental design that encourage standardization and interpretability. We utilized data elements for machine learning modeling building and refinement, from the electronic health record. We extracted standardized periodontal disease findings, diagnosis and treatment, attachment loss, gingivitis, and localized and generalized severe periodontitis. Clinician determined validated cases (n=5,180) were extracted with all EHR periodontal chart data, findings, diagnosis, procedure and digital radiographs for each case. 80% of the sampling data was used to apply the machine learning algorithms with interpretation of the radiographs for cases with good periodontal health as compared to patients with periodontitis (n=4,144). The remaining 20% of the extracted data was used to validate the system's ability to interpret radiographs without EHR elements (n=1,036).
Results: Of the 4,141 patient records, 96.6% were successfully utilized for machine learning algorithms. Data elements of 927,616 were successfully extracted and utilized. 14,494 of dental radiographs were utilized. The machine learning algorithm was able to consistently and accurately determine periodontal pocket depth with corresponding periodontal diagnosis and appropriateness of care.
Conclusions: By applying electronic health record derived periodontal findings, diagnosis and treatment in conjunction with dental radiographs, machine learning-based algorithms were able to accurately determine periodontal findings, diagnosis and appropriateness of care.