Table 3.
Resulting from replicating a prior study’s cross-validation, predicting treatment-resistant depression according to the Quick Inventory of Depressive Symptomatology, Clinician version (QID-C) scale, using data from Sequenced Treatment Alternatives to Relieve Depression.
GBDT: gradient boosting decision tree. Feature selection methods include clustering-χ^2(30 features) and elastic net (31 features). Results reported as Balanced Accuracy and area-under-curve (AUC). As the replicated study only reported one number for their results, we show the z-score of these against the distribution of our results from 100 runs of 10-fold cross-validation. Additional performance metrics and statistics are documented in S4 Table and S5 Table.
Table 4.
Performance of our predictive models when trained on the Sequenced Treatment Alternative to Relieve Depression (STAR*D) dataset, and externally evaluated on the Canadian Biomarker Integration Network in Depression (CAN-BIND-1) trial, predicting both response and remission according to the Quick Inventory of Depressive Symptomatology, Self Report Version (QIDS-SR) scale.
See Methods for our definition of these outcomes. No feature selection was used before running the models. Additional performance metrics and statistics are documented in S7 Table and S8 Table. GBDT: gradient boosting decision tree, AUC: area-under-curve.
Table 5.
Comparison of model performance with different targets and sets of features, using Random Forests.
Overlapping features are the 100 features in both Canadian Biomarker Integration Network in Depression’s CAN-BIND-1’s trial and Sequenced Treatment Alternatives to Relieve Depression (STAR*D), while Full uses all 480 features from STAR*D. Clustering-χ2 Selection (30 features) and Elastic Net Selection (31 features) refer to using these feature selection techniques as defined in Methods. Targets include antidepressant response, remission, or treatment-resistant depression (TRD), as defined in Methods. Models trained and evaluated using cross-validation (CV) on STAR*D, and we also report again the results of externally validating models on the CAN-BIND-1 dataset after being trained on STAR*D. We report balanced accuracy and area-under-curve (AUC). Additional performance metrics and statistics are documented in S10 Table and S11 Table. QIDS: Quick Inventory of Depressive Symptomatology, -SR: Self-Report, -C: Clinician.