IADR Abstract Archives
Machine Learning With Cross-Institutional Prediction for Orthodontic Tooth Extraction
Objectives: (1) To determine the performance of machine learning in prediction of extraction vs non-extraction for orthodontic patients from two university data sets
(2) To identify and compare rankings of most important features
Methods: Subjects consisted of 297 patients in the Graduate Clinic of Orthodontics at OSU for orthodontic treatment during a consecutive enrollment from 2017-2020. Input features (9 clinical, 11 cephalometric) were identified based on previous studies. Random forest (RF) model was trained using these feature sets on the sample population. The performance of each model was evaluated using measures including sensitivity, specificity, balanced accuracy, accuracy, positive predictive value (PPV) and negative predictive value (NPV). RF models trained by OSU data (OSU model) and previously obtained UNC data (UNC model) were used to predict on both OSU data and UNC data. Feature rank was calculated by RF to determine importance of each feature in extraction vs non-extraction decision.
Results: The sensitivity and specificity were 0.47 and 0.97, respectively, for OSU model and 0.32 and 0.94, respectively, for UNC model. Cross-prediction of data shows decrease in sensitivity for OSU model predicted on UNC, but similar sensitivity for UNC model predicted on OSU. Maxillary and mandibular crowding were the two most important features in both institutions.
Conclusions: The extraction vs non-extraction decision is a difficult challenge orthodontists face in clinical practice. AI expert system could provide new insight in addressing this long-standing debate. This study advanced the research in the field by applying it to multiple data centers in the US. Different philosophies/beliefs may exist in different institutions, but major factors are the same. Therefore, it is promising to use AI to predict results from different centers.
(2) To identify and compare rankings of most important features
Methods: Subjects consisted of 297 patients in the Graduate Clinic of Orthodontics at OSU for orthodontic treatment during a consecutive enrollment from 2017-2020. Input features (9 clinical, 11 cephalometric) were identified based on previous studies. Random forest (RF) model was trained using these feature sets on the sample population. The performance of each model was evaluated using measures including sensitivity, specificity, balanced accuracy, accuracy, positive predictive value (PPV) and negative predictive value (NPV). RF models trained by OSU data (OSU model) and previously obtained UNC data (UNC model) were used to predict on both OSU data and UNC data. Feature rank was calculated by RF to determine importance of each feature in extraction vs non-extraction decision.
Results: The sensitivity and specificity were 0.47 and 0.97, respectively, for OSU model and 0.32 and 0.94, respectively, for UNC model. Cross-prediction of data shows decrease in sensitivity for OSU model predicted on UNC, but similar sensitivity for UNC model predicted on OSU. Maxillary and mandibular crowding were the two most important features in both institutions.
Conclusions: The extraction vs non-extraction decision is a difficult challenge orthodontists face in clinical practice. AI expert system could provide new insight in addressing this long-standing debate. This study advanced the research in the field by applying it to multiple data centers in the US. Different philosophies/beliefs may exist in different institutions, but major factors are the same. Therefore, it is promising to use AI to predict results from different centers.