IADR Abstract Archives
TMJ Osteoarthritis Risk Assessment: Integrating Machine Learning and Slicer3D Deployment.
Objectives: Between 8 to 16% of the worldwide population experiences Temporomandibular Joint (TMJ) Osteoarthritis (OA). However, the risk factors driving its onset remain unidentified. This study aims to evaluate a novel machine learning and statistical models approach to predict risk factors of severe prognosis, authenticate this tailored prediction model in Python, and create a worldwide accessible tool in Slicer3D.
Methods: We employed the TMJ Prognosis model that leverages Nested Cross Validation, retaining the most effective out of 48 machine learning models. This is achieved by using 6 models for feature selection and 8 for prediction. The inner loop determines the optimal hyperparameter combination for the final model and the number of significant features. Initially, the dataset was divided into 10 parts for the outer loop. For each segment, training data was further split into 10 sub-segments in the inner loop. The testing set of the outer loop segments produced the model scores.
Results: Preliminary results combining NNET as the feature selection and Glmnet as the predictive model yielded the highest Accuracy (74.3%), AUC (76.3%) and F1 (68.1%) scores on the testing dataset. The top contributing features included: Headache, Restless sleep, Lower Back Pain, Age, saliva levels of Osteoprotegerin, Angiogenin, Chemokine Ligand 16 and Vascular-Endothelial-Growth-Factor, serum levels of Epithelial-Neutrophil-Activating peptide, Metalloproteinases Inhibitor 1, TNF-related activation-induced cytokine and Brain-Derived Neurotrophic Factor, Superior Joint Space, Articular fossa Grey Level non uniformity, long Run High Grey, Correlation and Bone Volume fraction, Condylar Bone Volume fraction, Short Run High Grey, Entropy, and Trabecular Bone Thickness.
Conclusions: The preliminary best performing model predicts patient prognosis, highlighting the most pertinent features for patient monitoring. Future steps will include the implementation of the Ensemble via Hierarchical Predictions through Nested cross validation method, which combines the top 3 performing models on the validation dataset with the six features selection methods.
Methods: We employed the TMJ Prognosis model that leverages Nested Cross Validation, retaining the most effective out of 48 machine learning models. This is achieved by using 6 models for feature selection and 8 for prediction. The inner loop determines the optimal hyperparameter combination for the final model and the number of significant features. Initially, the dataset was divided into 10 parts for the outer loop. For each segment, training data was further split into 10 sub-segments in the inner loop. The testing set of the outer loop segments produced the model scores.
Results: Preliminary results combining NNET as the feature selection and Glmnet as the predictive model yielded the highest Accuracy (74.3%), AUC (76.3%) and F1 (68.1%) scores on the testing dataset. The top contributing features included: Headache, Restless sleep, Lower Back Pain, Age, saliva levels of Osteoprotegerin, Angiogenin, Chemokine Ligand 16 and Vascular-Endothelial-Growth-Factor, serum levels of Epithelial-Neutrophil-Activating peptide, Metalloproteinases Inhibitor 1, TNF-related activation-induced cytokine and Brain-Derived Neurotrophic Factor, Superior Joint Space, Articular fossa Grey Level non uniformity, long Run High Grey, Correlation and Bone Volume fraction, Condylar Bone Volume fraction, Short Run High Grey, Entropy, and Trabecular Bone Thickness.
Conclusions: The preliminary best performing model predicts patient prognosis, highlighting the most pertinent features for patient monitoring. Future steps will include the implementation of the Ensemble via Hierarchical Predictions through Nested cross validation method, which combines the top 3 performing models on the validation dataset with the six features selection methods.