Final

• Correct the evaluation method. (Apply the training optimal threshold to test )

• Vote for the final prediction category

-One more random seed for repeated cross-validation has been added to TMJPI compared to the published paper. We do 11 times five-fold CV to avoid the edge case in which the half/half model votes for negative and positive.

-We calculated the final prediction score by binarizing the prediction probability with the optimal training threshold in each model. The most frequency category counted for the final prediction result.

Significantly different features

• Articular fossa and articular eminence were not significantly different between control and diseased groups and might not contribute to the diagnosis of the disease. These two set of features were excluded in this version.
• Only the 3D superior space and 6 condylar features (Energy,Entropy,Correlation,Cluster Prominence,HighGreyLevelRunEmphasis, ShortRunHighGreyLevelEmphasis) are significantly different between healthy and diseased groups.
• The model training includes all 5 clinical features (lastMonthDistressedHeadaches, lastMonthDistressedMuscleSoreness, verticalRangeUnassistedWOPain, verticalRangeUnassistedMax, cl_verticalRangeAssistedMax) and do feature selection from normal significant feature and biological features separately.

modified the model based on radio data without histogram matching

• Add Articular Eminence data into model training
• The model trained based on the updated data (no histogram matching on radio data)
• The order of features(column names) in the train*test should be identical
• The prediction result is the ensemble of 50 models (10 seeds * 5 Fold cross-validation)

Update the evaluation metrics

Generate "TMJPI_train_results.csv" to save all the evaluation metrics in training.

Including: arc, f1, sensitivity, specifity ,precision and accuracy

Updated feature selection

• NMIFS+ feature selection method now used for training the KRVFL+ model.
• Hyperparameter ranges updated for radiomic data that was histogram matched
• prediction binary no longer requires biological feature file

TMJPI

v1.0

Usage Instructions