Overview of data-driven MDD treatment response prediction studies. Relevant model performance metrics (BAC and R2) that were not reported by the studies but were possible to calculate from the reported values are included in parentheses. MDD - major depressive disorder, TRD - treatment-resistant depression, NDRI - norepinephrine-dopamine reuptake inhibitor, SSRI - selective serotonin reuptake inhibitor, SNRI - serotonin-norepinephrine reuptake inhibitor, rTMS - repetitive transcranial magnetic stimulation, tDCS - transcranial direct-current stimulation, ECT - electroconvulsive therapy, MADRS - Montgomery-Asberg Depression Rating Scale, HAMD - Hamilton Rating Scale for Depression, BDI - Beck Depression Inventory, QIDS-SR - Quick Inventory of Depressive Symptomatology—Self Report, ROI - region of interest, FC - functional connectivity, ICA - independent component analysis, SVM - support vector machine, LDA - linear discriminant analysis, CV - cross-validation, LOOCV - leave- one-out cross-validation, LOSOCV - leave-one-site-out cross-validation, Acc - accuracy, SE - sensitivity, SP - specificity, BAC - balanced accuracy, RMSE - root mean square error, dmPCF - dorsomedial prefrontal cortex, ACC - anterior cingulate cortex.
| Reference . | Subjects . | Treatment . | Modality . | Outcome defintion . | Features . | Model . | Validation . | Performance . |
|---|---|---|---|---|---|---|---|---|
| Jaworska et al. (2019) | 51 MDD | NDRI (bupropion), SSRI (escitalopram), or combination of both | rsEEG | Response: ≥50% ↓ in MADRS | Demographics, baseline, & Week 1 clinical data, EEG power features, current source density | Random forest | 10-fold CV | Acc 88% |
| SE 77% | ||||||||
| SP 99% | ||||||||
| Zhdanov et al. (2020) | 122 MDD | SSRI (escitalopram) | rsEEG | Response: ≥50% ↓ in MADRS | Electrode-level & source-level spectral features, multiscale entropy-based & microstate-based features | SVM | Leave-one-site-out CV (LOSOCV) | BAC 79% |
| SE 67% | ||||||||
| SP 91% | ||||||||
| Khodayari-Rostamabad et al. (2010) | 22 MDD | SSRI (mainly sertraline) | rsEEG | Response: ≥25% ↓ in HAMD-17 | Spectral coherence, mutual information between electrode pairs, absolute & relative power spectral density | Kernel partial least squares regression | Nested CV | Acc 87% |
| SE 88% | ||||||||
| SP 86% | ||||||||
| Khodayari-Rostamabad et al. (2013) | 22 TRD | SSRI (sertraline, citalopram, fluvoxamine, or paroxetine) | rsEEG | Response: ≥30% ↓ in HAMD-17 | Power spectral density, squared spectral coherence, mutual information, left-to-right hemispheres, & anterior/posterior power ratio | Mixture of factor analysis | k-fold CV | Acc 88% |
| SE 95% | ||||||||
| SP 81% | ||||||||
| Rabinoff et al. (2011) | 25 MDD | SSRI (fluoxetine) or SNRI (venlafaxine) | rsEEG | Response: HAMD-17 ≤ 10 | Absolute & relative power, cordance features | Classification and regression trees (CART) | LOOCV | BAC 93% |
| SE 85% | ||||||||
| SP 100% | ||||||||
| Shahabi, Shalbaf, and Maghsoudi (2021) | 30 MDD | SSRI (type not specified) | rsEEG | Response: ≥50% ↓ in BDI | 3D images constructed from EEG signal | Convolutional neural networks | 10-fold CV | Acc 97% |
| SE 96% | ||||||||
| SP 97% | ||||||||
| W. Wu et al. (2020) | 109 MDD (sertraline), 119 MDD (placebo) | SSRI (sertraline) | rsEEG | Δ in HAMD-17 | Theta, alpha, beta, gamma band power of latent signal | Linear regression | 10-fold CV | (R2 = 0.36) |
| r = 0.60 | ||||||||
| RMSE = 5.68 | ||||||||
| p = 2.88 × 10−11 | ||||||||
| Rajpurkar et al. (2020) | 518 MDD | SSRI (escitalopram, sertraline) or SNRI (venlafaxine) | rsEEG | Δ in HAMD-21 (individual symptoms) | Absolute & relative power of delta, theta, alpha, beta, & gamma frequency bands in frontal & occipital regions | Gradient-boosted decision trees (GBDT) | 5-fold stratified CV | Concordance index of ≥0.8 on 12 out of 21 symptoms |
| R2 0.3–0.7 | ||||||||
| Khodayari-Rostamabad, Reilly, Hasey, de Bruin, and MacCrimmon (2011) | 27 TRD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 | Anterior/posterior power ratios at various frequencies | Mixture of factor analysis | k-fold CV | Acc 80% |
| (BAC 81%) | ||||||||
| SE 78% | ||||||||
| SP 83% | ||||||||
| N. Bailey et al. (2019) | 50 TRD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 | Mood features, theta & alpha power & connectivity, frontal theta cordance & alpha peak frequency | SVM | 5-fold CV | BAC 86% |
| SE 84% | ||||||||
| SP 89% | ||||||||
| Hasanzadeh et al. (2019) | 46 MDD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 or BDI-II | Nonlinear, power spectral density, bispectrum, frontal & prefrontal cordance | k-nearest neighbors | LOOCV | BAC 91% |
| SE 87% | ||||||||
| SP 96% | ||||||||
| Al-Kaysi et al. (2017) | 10 MDD | tDCS | rsEEG | Response: ≥50% ↓ in MADRS | Power spectral density in delta, theta, alpha, beta, & gamma frequency bands | SVM, LDA, extreme learning machine | LOOCV | Mood Labels Channels FC4-AF8: Acc 76%; Cognition Labels Channels CPz-CP2: Acc 92% |
| Tian et al. (2020) | 106 MDD | SSRI (escitalopram) | rsfMRI | Response: ≥50% ↓ in HAMD-17 | Multilayer modularity framework applied to the whole brain to obtain measures of functional integration & segregation among 95 ROIs | SVM | Leave-one-site-out CV (LOSOCV) | BAC 71% |
| Klöbl et al. (2020) | 29 MDD | SSRI (escitalopram) | rsfMRI | Δ in HAMD-17; Response: ≥50% ↓ in HAMD-17; Remission: HAMD-17 ≤ 7 | Whole-brain FC | Linear regression | k-fold CV | HAMD-sum: r = 0.51 |
| Response: BAC 60%, AUC 68% | ||||||||
| Remission: BAC 68%, AUC 73% | ||||||||
| Chin Fatt et al. (2020) | 132 MDD (sertraline), 132 MDD (placebo) | SSRI (sertraline) | rsfMRI | Δ in HAMD-17 | Cortical & subcortical seed-based FC | Linear mixed model | LOOCV | (R2 = 0.05–0.13) |
| r = 0.22–0.36 | ||||||||
| Korgaonkar et al. (2020) | 163 MDD | SSRI (escitalopram, sertraline) or SNRI (venlafaxine) | rsfMRI | Remission: HAMD-17 ≤ 7 | Whole-brain network intrinsic FC | Logistic regression | Hold-out test set | Average connectivity measures: |
| Acc 69% | ||||||||
| (BAC 67%) | ||||||||
| SE 58% | ||||||||
| SP 76% | ||||||||
| Individual network connectivity: | ||||||||
| Acc 69% | ||||||||
| (BAC 68%) | ||||||||
| SE 63% | ||||||||
| SP 72% | ||||||||
| Nemati et al. (2020) | 99 MDD (sertraline), 103 MDD (placebo), & 19 MDD (ketamine), 19 MDD (active control), 18 MDD (inactive control) | SSRI (sertraline), ketamine | rsfMRI | Δ in HAMD-17 | Network restricted connectivity | Network restricted strength predictive (linear) model | 10-fold CV | Sertraline (vs. placebo): r = 0.27 (R2 = 0.07), p = 0.003; Ketamine (vs. active placebo): r = 0.57 (R2 = 0.32), p = 0.0002 |
| Fan et al. (2020) | 97 MDD (sertraline), 103 MDD (placebo) | SSRI (sertraline) | rsfMRI | Δ in HAMD-17 (%) | Network restricted connectivity | Network restricted strength predictive (linear) model | 10-fold CV | Response to sertaline or placebo: (R2 = 0.04) r = 0.19, p = 0.03 |
| Ju et al. (2020) | 108 MDD | Various drugs; primarily: paroxetine, other SSRIs, sedative hypnotics, NDRI (bupropion) | rsEEG | Δ in HAMD-24 | Whole-brain FC matrices | Connectome-based predictive modeling | LOOCV | r = 0.43 (R2 = 0.19), p = 2.73 × 10−6 |
| Kong et al. (2021) | 82 MDD | Antidepressants (type not specified) | rsfMRI | Response: >50% ↓ in HAMD-21 | Dynamic functional networks | Spatiotemporal graph convolutional network | 10-fold CV | Acc 90% |
| (BAC 89%) | ||||||||
| SE 85% | ||||||||
| SP 93% | ||||||||
| Drysdale et al. (2017) | 154 MDD | rTMS | rsfMRI | Response: ≥50% ↓ in HAMD-17 | Whole-brain FC matrices & biotype diagnosis | SVM | LOOCV | Only FC feature: Acc 78.3% |
| FC features & biotype diagnosis: Acc 89.6% | ||||||||
| Van Waarde et al. (2015) | 45 severe/TRD | ECT | rsfMRI | Remission: MADRS score ≤ 10 | Standard group ICA extracted 25 rs-networks. Each network was used to train a classifier | SVM | LOOCV | Two rs-networks had significant accuracy: dmPFC: BAC 85%; SE 84%; SP 85%; ACC: BAC 78%; SE 80%; SP 75% |
| Leaver et al. (2018) | 46 TRD | ECT | rsfMRI, sMRI, arterial spin labeled fMRI | Response: average % improvement in HAMD-17, MADRS, and QIDS-SR. Split point was 42.2% reduction. | Mean voxelwise cerebral blood flow, regional homogeneity, fractional amplitude of low-frequency fluctuations, gray matter volume | SVM | Nested CV | BAC 58–68% |
| SE 54–64% | ||||||||
| SP 55–74% | ||||||||
| Sun et al. (2020) | 122 MDD or bipolar disorder | ECT | rsfMRI | Δ in HAMD-17; Remission: HAMD-17 score < 7 | Negatively & positively correlated FC networks based on whole-brain rsFC | Linear regression | LOOCV | Negative FC networks: |
| r = 0.51 | ||||||||
| (R2 = 0.26) | ||||||||
| Acc 76% | ||||||||
| (BAC 72%) | ||||||||
| SE 51% | ||||||||
| SP 92% |
| Reference . | Subjects . | Treatment . | Modality . | Outcome defintion . | Features . | Model . | Validation . | Performance . |
|---|---|---|---|---|---|---|---|---|
| Jaworska et al. (2019) | 51 MDD | NDRI (bupropion), SSRI (escitalopram), or combination of both | rsEEG | Response: ≥50% ↓ in MADRS | Demographics, baseline, & Week 1 clinical data, EEG power features, current source density | Random forest | 10-fold CV | Acc 88% |
| SE 77% | ||||||||
| SP 99% | ||||||||
| Zhdanov et al. (2020) | 122 MDD | SSRI (escitalopram) | rsEEG | Response: ≥50% ↓ in MADRS | Electrode-level & source-level spectral features, multiscale entropy-based & microstate-based features | SVM | Leave-one-site-out CV (LOSOCV) | BAC 79% |
| SE 67% | ||||||||
| SP 91% | ||||||||
| Khodayari-Rostamabad et al. (2010) | 22 MDD | SSRI (mainly sertraline) | rsEEG | Response: ≥25% ↓ in HAMD-17 | Spectral coherence, mutual information between electrode pairs, absolute & relative power spectral density | Kernel partial least squares regression | Nested CV | Acc 87% |
| SE 88% | ||||||||
| SP 86% | ||||||||
| Khodayari-Rostamabad et al. (2013) | 22 TRD | SSRI (sertraline, citalopram, fluvoxamine, or paroxetine) | rsEEG | Response: ≥30% ↓ in HAMD-17 | Power spectral density, squared spectral coherence, mutual information, left-to-right hemispheres, & anterior/posterior power ratio | Mixture of factor analysis | k-fold CV | Acc 88% |
| SE 95% | ||||||||
| SP 81% | ||||||||
| Rabinoff et al. (2011) | 25 MDD | SSRI (fluoxetine) or SNRI (venlafaxine) | rsEEG | Response: HAMD-17 ≤ 10 | Absolute & relative power, cordance features | Classification and regression trees (CART) | LOOCV | BAC 93% |
| SE 85% | ||||||||
| SP 100% | ||||||||
| Shahabi, Shalbaf, and Maghsoudi (2021) | 30 MDD | SSRI (type not specified) | rsEEG | Response: ≥50% ↓ in BDI | 3D images constructed from EEG signal | Convolutional neural networks | 10-fold CV | Acc 97% |
| SE 96% | ||||||||
| SP 97% | ||||||||
| W. Wu et al. (2020) | 109 MDD (sertraline), 119 MDD (placebo) | SSRI (sertraline) | rsEEG | Δ in HAMD-17 | Theta, alpha, beta, gamma band power of latent signal | Linear regression | 10-fold CV | (R2 = 0.36) |
| r = 0.60 | ||||||||
| RMSE = 5.68 | ||||||||
| p = 2.88 × 10−11 | ||||||||
| Rajpurkar et al. (2020) | 518 MDD | SSRI (escitalopram, sertraline) or SNRI (venlafaxine) | rsEEG | Δ in HAMD-21 (individual symptoms) | Absolute & relative power of delta, theta, alpha, beta, & gamma frequency bands in frontal & occipital regions | Gradient-boosted decision trees (GBDT) | 5-fold stratified CV | Concordance index of ≥0.8 on 12 out of 21 symptoms |
| R2 0.3–0.7 | ||||||||
| Khodayari-Rostamabad, Reilly, Hasey, de Bruin, and MacCrimmon (2011) | 27 TRD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 | Anterior/posterior power ratios at various frequencies | Mixture of factor analysis | k-fold CV | Acc 80% |
| (BAC 81%) | ||||||||
| SE 78% | ||||||||
| SP 83% | ||||||||
| N. Bailey et al. (2019) | 50 TRD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 | Mood features, theta & alpha power & connectivity, frontal theta cordance & alpha peak frequency | SVM | 5-fold CV | BAC 86% |
| SE 84% | ||||||||
| SP 89% | ||||||||
| Hasanzadeh et al. (2019) | 46 MDD | rTMS | rsEEG | Response: ≥50% ↓ in HAMD-17 or BDI-II | Nonlinear, power spectral density, bispectrum, frontal & prefrontal cordance | k-nearest neighbors | LOOCV | BAC 91% |
| SE 87% | ||||||||
| SP 96% | ||||||||
| Al-Kaysi et al. (2017) | 10 MDD | tDCS | rsEEG | Response: ≥50% ↓ in MADRS | Power spectral density in delta, theta, alpha, beta, & gamma frequency bands | SVM, LDA, extreme learning machine | LOOCV | Mood Labels Channels FC4-AF8: Acc 76%; Cognition Labels Channels CPz-CP2: Acc 92% |
| Tian et al. (2020) | 106 MDD | SSRI (escitalopram) | rsfMRI | Response: ≥50% ↓ in HAMD-17 | Multilayer modularity framework applied to the whole brain to obtain measures of functional integration & segregation among 95 ROIs | SVM | Leave-one-site-out CV (LOSOCV) | BAC 71% |
| Klöbl et al. (2020) | 29 MDD | SSRI (escitalopram) | rsfMRI | Δ in HAMD-17; Response: ≥50% ↓ in HAMD-17; Remission: HAMD-17 ≤ 7 | Whole-brain FC | Linear regression | k-fold CV | HAMD-sum: r = 0.51 |
| Response: BAC 60%, AUC 68% | ||||||||
| Remission: BAC 68%, AUC 73% | ||||||||
| Chin Fatt et al. (2020) | 132 MDD (sertraline), 132 MDD (placebo) | SSRI (sertraline) | rsfMRI | Δ in HAMD-17 | Cortical & subcortical seed-based FC | Linear mixed model | LOOCV | (R2 = 0.05–0.13) |
| r = 0.22–0.36 | ||||||||
| Korgaonkar et al. (2020) | 163 MDD | SSRI (escitalopram, sertraline) or SNRI (venlafaxine) | rsfMRI | Remission: HAMD-17 ≤ 7 | Whole-brain network intrinsic FC | Logistic regression | Hold-out test set | Average connectivity measures: |
| Acc 69% | ||||||||
| (BAC 67%) | ||||||||
| SE 58% | ||||||||
| SP 76% | ||||||||
| Individual network connectivity: | ||||||||
| Acc 69% | ||||||||
| (BAC 68%) | ||||||||
| SE 63% | ||||||||
| SP 72% | ||||||||
| Nemati et al. (2020) | 99 MDD (sertraline), 103 MDD (placebo), & 19 MDD (ketamine), 19 MDD (active control), 18 MDD (inactive control) | SSRI (sertraline), ketamine | rsfMRI | Δ in HAMD-17 | Network restricted connectivity | Network restricted strength predictive (linear) model | 10-fold CV | Sertraline (vs. placebo): r = 0.27 (R2 = 0.07), p = 0.003; Ketamine (vs. active placebo): r = 0.57 (R2 = 0.32), p = 0.0002 |
| Fan et al. (2020) | 97 MDD (sertraline), 103 MDD (placebo) | SSRI (sertraline) | rsfMRI | Δ in HAMD-17 (%) | Network restricted connectivity | Network restricted strength predictive (linear) model | 10-fold CV | Response to sertaline or placebo: (R2 = 0.04) r = 0.19, p = 0.03 |
| Ju et al. (2020) | 108 MDD | Various drugs; primarily: paroxetine, other SSRIs, sedative hypnotics, NDRI (bupropion) | rsEEG | Δ in HAMD-24 | Whole-brain FC matrices | Connectome-based predictive modeling | LOOCV | r = 0.43 (R2 = 0.19), p = 2.73 × 10−6 |
| Kong et al. (2021) | 82 MDD | Antidepressants (type not specified) | rsfMRI | Response: >50% ↓ in HAMD-21 | Dynamic functional networks | Spatiotemporal graph convolutional network | 10-fold CV | Acc 90% |
| (BAC 89%) | ||||||||
| SE 85% | ||||||||
| SP 93% | ||||||||
| Drysdale et al. (2017) | 154 MDD | rTMS | rsfMRI | Response: ≥50% ↓ in HAMD-17 | Whole-brain FC matrices & biotype diagnosis | SVM | LOOCV | Only FC feature: Acc 78.3% |
| FC features & biotype diagnosis: Acc 89.6% | ||||||||
| Van Waarde et al. (2015) | 45 severe/TRD | ECT | rsfMRI | Remission: MADRS score ≤ 10 | Standard group ICA extracted 25 rs-networks. Each network was used to train a classifier | SVM | LOOCV | Two rs-networks had significant accuracy: dmPFC: BAC 85%; SE 84%; SP 85%; ACC: BAC 78%; SE 80%; SP 75% |
| Leaver et al. (2018) | 46 TRD | ECT | rsfMRI, sMRI, arterial spin labeled fMRI | Response: average % improvement in HAMD-17, MADRS, and QIDS-SR. Split point was 42.2% reduction. | Mean voxelwise cerebral blood flow, regional homogeneity, fractional amplitude of low-frequency fluctuations, gray matter volume | SVM | Nested CV | BAC 58–68% |
| SE 54–64% | ||||||||
| SP 55–74% | ||||||||
| Sun et al. (2020) | 122 MDD or bipolar disorder | ECT | rsfMRI | Δ in HAMD-17; Remission: HAMD-17 score < 7 | Negatively & positively correlated FC networks based on whole-brain rsFC | Linear regression | LOOCV | Negative FC networks: |
| r = 0.51 | ||||||||
| (R2 = 0.26) | ||||||||
| Acc 76% | ||||||||
| (BAC 72%) | ||||||||
| SE 51% | ||||||||
| SP 92% |