Fig 1.
After extracting DFs, we combined CF and DF to make three different combinations of features. After feature set generation, we fed each feature set into nine different classifiers to predict PD-CD based on two cognitive scales and four timeframes. CF: Clinical feature, DF: Deep features, RFE: Recursive Feature Elimination, ADBC: AdaBoost Classifier, BGC: Bagging Classifier, SVC: Support Vector Classifier, KNN: K-Nearest Neighbors Classifier, RFC: Random Forest Classifier, GBC: Gradient Boosting Classifier, MLP: Multi-Layer Perceptron, DTC: Decision Tree Classifier, LRC: Logistic Regression Classifier, CF: Clinical Features, DF: Deep Features.
Fig 2.
Heatmap for cross-validation AUC scores.
Showing the superiority of MoCA performance compared to MDS-UPDRS-I and also the high performance of year 4 in both scales. ADBC: AdaBoost Classifier, BGC: Bagging Classifier, SVC: Support Vector Classifier, KNN: K-Nearest Neighbors Classifier, RFC: Random Forest Classifier, GBC: Gradient Boosting Classifier, MLP: Multi-Layer Perceptron, DTC: Decision Tree Classifier, LRC: Logistic Regression Classifier, CF: Clinical Features, DF: Deep Features.
Fig 3.
Showcasing the superior results for both the MoCA and MDS-UPDRS-I across all three feature sets. The peak AUC performance for the MoCA was secured through the deployment of DF for forecasting the MoCA score in year 4. Similarly, the optimal AUC for the MDS-UPDRS-I was also obtained by employing DF to estimate the MDS-UPDRS-I score in the same year. CF: Clinical feature, DF: Deep features.