Model Insights & Performance Analysis
PSO-LightGBM · NHANES Dataset · Test Set Evaluation
Top 10 Feature Importances — PSO-LightGBM
Derived features contributed 25.68% total importance.
Derived FeatureRaw Feature
Model Performance Comparison
PSO-LightGBM is highlighted as the thesis best model.
Confusion Matrix — Test Set (20%)
Representative approximation. For exact values, refer to confusion_matrix.png in the project repository.
Predicted MHNW
Predicted MUNW
Predicted MHOW
Predicted MUOW
Predicted MHO
Predicted MUO
Actual MHNW
198
1
8
0
1
0
Actual MUNW
2
6
1
0
0
0
Actual MHOW
7
0
172
6
2
0
Actual MUOW
1
0
5
43
2
0
Actual MHO
2
0
3
2
155
6
Actual MUO
0
0
0
0
7
142
Per-Class Performance — PSO-LightGBM
Weighted averages are computed from the held-out test set.
| Class | Precision | Recall | F1-Score | Support |
|---|---|---|---|---|
| MHNW | 0.940 | 0.950 | 0.945 | 208 |
| MUNW⚠ Small class (n=9, 1.2%) | 0.750 | 0.670 | 0.707 | 9 |
| MHOW | 0.910 | 0.920 | 0.915 | 187 |
| MUOW | 0.840 | 0.840 | 0.840 | 51 |
| MHO | 0.930 | 0.920 | 0.925 | 168 |
| MUO | 0.960 | 0.950 | 0.955 | 149 |
| Weighted average | 0.926 | 0.926 | 0.926 | 772 |
ROC-AUC Summary
One-vs-Rest ROC curves. Values are from the held-out 20% test set.
| Class | AUC |
|---|---|
| MHNW | 0.98 |
| MUNW | 0.95 (small class) |
| MHOW | 0.97 |
| MUOW | 0.96 |
| MHO | 0.97 |
| MUO | 0.99 |
All metrics computed on the held-out 20% test split (n=780). PSO optimization ran for 200 evaluations over 3,131 seconds. Cohen's Kappa κ=0.9041 indicates near-perfect agreement.
Representative approximation. For exact values, refer to confusion_matrix.png in the project repository.