Table 2.

Previous studies predicting depression treatment response.

Author	Patient sample	Features [Imaging modality]	Feature reduction method	Cross-validation method	Machine learning method	Results (Note: highest accuracies presented)
Costafreda et al. (2009)	– 9 responders – 9 non-responders	– Smoothed gray matter voxel-based intensity values [T1-weighted]	– Voxel based morphometry – Filter method using ANOVA	– Leave-one-out cross-validation	– Support vector machines	– Accuracya: 88.9% – Sensitivityb: 88.9% – Specificityc: 88.9%
Liu et al. (2012)	– 17 responders – 18 non-responders	– Gray and white matter smoothed voxel-based intensity values [T1-weighted]	– Multivariate pattern analysis – Searchlight algorithm – Principal component analysis	– Leave-one-out cross-validation	– Support vector machines (linear kernel)	– Accuracya: 82.9%
Marquand et al. (2008)	– 9 responders – 9 non-responders	– Smoothed whole brain voxel-based blood oxygen level dependent response during a verbal working memory fMRI task [fMRI]	– Principal component analysis	– Leave-one-out cross-validation	– Support vector machines (linear kernel)	– Accuracya: 69% – Sensitivityb: 85% – Specificityc: 52%
Nouretdinov et al. (2011)	– 9 responders – 9 non-responders	– Smoothed voxel-based intensity values [T1-weighted]	– n/a	– Leave-one-out cross-validation	– Support vector machines (linear kernel) with general probabilistic classification method (transductive conformal predictor)	– Accuracya: 83.3% – Sensitivityb: 77.8% – Specificityc: 88.9%

^aOverall classification accuracy.

^bPercent responders identified.

^cPercent non-responders identified.