TABLE 6.
Best accuracies and related ROC AUC measures for case‐control classification (on cross‐validation) for brain moprhometric and white matter integrity features in cMDD‐UKB, rMDD‐STR, pMDD‐UKB‐CIDI, and pMDD‐UKB‐ICD samples
| Data set | Feature domain | Sample size | Classification approach | Classification accuracy (sensitivity/specificity) | ROC AUC |
|---|---|---|---|---|---|
| cMDD‐UKB | Cortical thickness features | 1,470 |
PLR classifier ‐ Hyperparameter grid search ‐ Embedded feature selection |
52.80% (52.66/52.92%) |
0.540 |
| Combined FA and MD features | 2,870 |
SVM classifier ‐ No hyperparameter optim. ‐ Sequential feat. Elimination |
53.73% (51.08/56.37%) |
0.549 | |
| rMDD‐STR | Combined brain morphometric features | 296 |
Decision tree classifier ‐ Hyperparameter grid search ‐ Filter feature selection |
57.48% (52.57/62.35%) |
0.572 |
| MD features | 404 |
SVM classifier ‐ Hyperparameter grid search ‐ No feature selection |
55.54% (59.16/51.92%) |
0.560 | |
| pMDD‐UKB‐CIDI | Cortical thickness features | 3,330 |
SVM classifier ‐ No hyperparameter optim. ‐ No feature selection |
53.63% (53.72/53.54%) |
0.532 |
| Combined FA and MD features | 6,836 |
SVM classifier ‐ 5‐fold inner/outer CV ‐ No hyperparameter optim. ‐ Sequential feat. Elimination |
52.68% (53.63/51.73%) |
0.531 | |
| pMDD‐UKB‐ICD | Combined brain morphometric features | 280 |
PLR classifier ‐ Hyperparameter grid search ‐ Embedded feature selection |
60.29% (61.86/58.71%) |
0.645 |
| MD features | 578 |
SVM classifier ‐ No hyperparameter optim. ‐ Filter feature selection |
56.18% (68.56/43.83%) |
0.566 |
Note: Combined brain morphometric features included cortical thickness, surface area, cortical and subcortical volume measures. Nested 10‐fold outer and 10‐fold inner cross‐validation was performed in all analyses, except where otherwise specified. Top two accuracies are in italics.
Abbreviations: cMDD‐UKB, sample with probable current MDD in UK Biobank cohort; FA, fractional anisotropy; MD, mean diffusivity; PLR, penalised logistic regression; pMDD‐UKB‐CIDI, sample with lifetime MDD based on the online Composite International Diagnostic Interview (CIDI) criteria in UK Biobank cohort; pMDD‐UKB‐ICD, sample with lifetime MDD based on the ICD criteria and hospital records in UK Biobank cohort; rMDD‐STR, sample with remitted MDD in STRADL cohort; ROC AUC, receiver operating characteristic area under the curve; SVM, support vector machine.