Page 125 - DP-2-3

Design+ ML for predicting Alzheimer’s progression

Figure 4. Feature importance using permutation method

LIMMTOTAL. While MH9ENDO ranked next in 5.4. Performance evaluation
importance, its contribution during the imputation As shown in Table 2, the performance evaluation metrics
process was relatively less significant compared to the indicate that the tuned RF model with selected features
other features. This comprehensive analysis underscores outperformed the other models. The best hyperparameters
the pivotal role of these features in predicting the target included “n_estimators” = 100, “min_samples_split” = 15,
outcome, thereby guiding subsequent steps in the analysis. “min_samples_leaf” = 1, and “max_depth” = 50—identified

5.2. Feature selection through randomized search with five-fold cross-validation
over 100 iterations. Furthermore, the tuned RF model with
The RF-based feature selection process identified four selected features demonstrated exceptional performance
key features as crucial for predicting the output class, across multiple evaluation metrics.
DXCURREN: CDGLOBAL, MMSCORE, LIMMTOTAL,
and LDELTOTAL. These features demonstrated In Class 0 (HC), the model exhibited high precision
significant importance in accurately predicting the target (97%) and recall (93%), ensuring accurate identification
outcome. Additionally, the feature importance analysis of HCs. For Class 1 (MCI), while precision was moderate
revealed MH9ENDO as an additional feature, though its (69%), the model displayed commendable recall (89%),
contribution was relatively minor compared to the others. which is crucial for identifying MCI instances. Class 2
Implementing this feature selection approach supported (AD) demonstrated balanced precision (95%) and recall
decision-making by excluding MH9ENDO from the final (78%), essential for accurately identifying AD cases.
feature set. The overall accuracy of 90% underscores the model’s
efficiency, with both macro-average (precision: 0.87;
5.3. Class balancing recall: 0.87; F1-score: 0.86) and weighted-average metrics
In the dataset exhibiting class imbalance, the sample (precision: 0.91; recall: 0.90; F1-score: 0.90) confirming its
distribution was skewed, with 609 samples for HC, 144 for consistency and superior performance across all classes.
MCI, and 105 for AD. By generating additional synthetic This comprehensive evaluation highlights the effectiveness
samples using SMOTE, each class was balanced to contain of the tuned RF model in accurately distinguishing
490 samples. between different diagnostic categories. The Appendix

Volume 2 Issue 3 (2025) 7 doi: 10.36922/DP025270031