Table 1 Machine learning studies predicting treatment response using EEG in major depressive disorder (a summary of sample size, treatment outcomes, machine learning algorithms, and performance metrics).
First author, year | Intervention | Outcome | Machine learning model | Accuracy | Other measures | |
|---|---|---|---|---|---|---|
Studies predicting response to neurostimulation therapy | ||||||
Bailey [26] | 39 patients with treatment-resistant depression | 3 weeks (15 sessions) unilateral left 10 Hz rTMS | Responders (≥50% decrease in HAM-D after 5–8 weeks of rTMS) vs. Non-responders | Linear SVM | 91% | Sensitivity: 91% Specificity: 92% F1 score: 0.93 |
Bailey [26] | 32 patients with treatment-resistant depression | 3 weeks (15 sessions) unilateral left 10 Hz rTMS | Responders (≥50% decrease in HAM-D after 5–8 weeks of rTMS) vs. Non-responders | Linear SVM | 86.66% | Sensitivity: 84% Specificity: 89% |
Corlier [28] | 109 patients with MDD | 3 weeks (15 sessions) of 10 Hz left DLPFC rTMS (68 subjects received unilateral left treatment, 41 were changed to sequential bilateral treatment—10 Hz left DLPFC, 1 Hz right DLPFC) | Responders (≥40% decrease in IDS-30 scores from baseline to treatment 30) vs. Non-responders | Elastic Net | 61.8–79.2% (Best performance observed with alpha band frequency and IDS-30 percent change score) | AUC: 0.52–0.77 Specificity: 70.9–82.7% Sensitivity: 34.8–75.7% PPV: 58.2–79.7% NPV: 63.8–82.2% |
Erguzel [29] | 147 patients with treatment-resistant depression | 18 sessions of 25 Hz left PFC rTMS | Responders (≥50% decrease in HAM-D scores after 3 weeks of treatment) vs. Non-responders | BPNN | 89.12% | Sensitivity: 94.44% AUC: 0.904 |
Erguzel [30] | 55 patients with MDD | 18 sessions of 25 Hz left PFC rTMS | Responders (≥50% decrease in HAM-D scores after 3 weeks of treatment) vs. Non-responders | ANN | 89.09% | Sensitivity: 86.67–93.33% Specificity: 80–84% AUC: 0.686–0.909 Best model (6-fold CV) Sensitivity: 93.3% Specificity: 84.0% AUC: 0.909 |
Erguzel [31] | 147 patients with treatment-resistant depression | 20 sessions of adjunctive 25 Hz left PFC rTMS | Responders (≥50% decrease in HAM-D scores after 20 sessions of rTMS) vs. Non-responders | ANN SVM DT | Accuracy: 78.3–86.4% Best performance using SVM Balanced Accuracy: 54.71–75.42% | Sensitivity: 60.41–68.62% Specificity: 49.01–82.22% |
Hasanzadeh [33] | 46 patients with MDD | 5-sessions of 10 Hz left DLPFC rTMS | Responders (≥50% decrease in BDI-II or HAM-D scores from baseline) vs. Non-responders Remission (Remission defined as BDI ≤ 8 or HAM-D ≤ 9) vs. Non-remission | kNN | 76.1–91.3% best performance with power spectral features | Sensitivity: 69.6–87% Specificity: 82.6–95.7% |
Studies predicting response to pharmacological treatment | ||||||
Cao [34] | 37 patients with treatment-resistant depression | Patients randomized to one of three groups (1:1:1): 0.5 mg/kg ketamine 0.2 mg/kg ketamine Normal saline | Responders (≥45% reduction in HAM-D score from baseline to 240 min posttreatment) vs. Non-responders | LDA NMSC kNN PARZEN PERLC DRBMC SVM Radial kernel | 78.4% Best performance using SVM with a radial kernel | Sensitivity: 79.3% Specificity: 84.2% Recall: 78.5% Precision: 87.0% F1 score: 52.6% |
Cook [35] | 180 patients with MDD | 8-week trial of escitalopram (10 mg) or bupropion (150 mg) (1-week single-blind escitalopram followed by 7 weeks double-blind trial) | Remission (≤7 HDRS at week 8) vs. Non-remission | LDA | 64.4% | Sensitivity: 74.3% Specificity: 55.3% PPV: 60.5% NPV: 70.0% AUC: 0.635 |
de la Salle [36] | 47 patients with MDD | 12-week double-blinded trial of: (1) escitalopram+ bupropion (2) escitalopram+ placebo (3) bupropion+placebo | Responders (≥50% reduction in MADRS scores from baseline to posttreatment) vs. Non-responders Remitters (≤10 MADRS at post-treatment) vs. Non-responders | LR | Response: Change in PF Cordance: 81% Change in MRF Cordance: 74% Remission: Change in PF Cordance: 70% Change in MRF Cordance: 51% | Response (ΔPF): AUC: 0.85 Sensitivity: 70% Specificity: 85% PPV: 0.95 NPV: 0.76 Remission (ΔPF): AUC: 0.66 Sensitivity: 65% Specificity: 74% PPV: 65% NPV: 74% Response (ΔMRF): AUC: 0.80 Sensitivity: 70% Specificity: 95% PPV: 95% NPV: 76% Remission (ΔMRF): AUC: 0.59 Sensitivity: 93% Specificity: 31% PPV: 39% NPV: 91% |
Jaworska [37] | 51 patients with MDD | 12-week double-blinded trial of: (1) escitalopram+bupropion (2) escitalopram+placebo (3) bupropion+placebo | Responders (≥50% reduction in MADRS scores from baseline to posttreatment) vs. Non-responders | RF SVM AdaBoost CART MLP GNB | 88% Combined model, accuracy of each individual model not reported | AUC: 0.716-0.901 Highest AUC observed in Random Forest Model Combined model Sensitivity = 77% Specificity = 99% PPV = 99 NPV = 81 |
Mumtaz [38] | 34 patients with MDD | Open-label trial of an SSRI | Responders (Responders defined as ≥50% improvement in pre- vs. post-treatment BDI-II scores) vs. Non-responders | LR | 87.5% | Sensitivity: 95% Specificity: 80% |
Rajpurkar [39] | 518 patients with MDD | Patients randomized in a 1:1:1: ratio to escitalopram, sertraline, or extended-release venlafaxine for 8 weeks | Regression model (Continuous improvement in individual symptoms, defined as the difference in score for each of the symptoms on the HAM-D from baseline to week 8) | GBM | R2 = 0.375–0.551 Best model observed using EEG and baseline symptom features | 95% CI: 0.473–0.639 Used C-index to assess performance (probability that the algorithm will correctly identify, given 2 random patients with different improvement levels, which patient showed greater improvement |
Wu [40] | 309 patients with MDD | 8-week course of sertraline or placebo | Regression model (Pre- minus post-treatment difference in HAMD17 scores, with missing endpoint values, imputed to maintain an intent-to-treat framework.) | SELSER Algorithm developed in the current study | R2 = 0.60 Sertraline R2 = 0.41 Placebo | NA |
Zhdanov [41] | 122 patients with MDD | 8-weeks of open-label escitalopram (10–20 mg) treatment | Responders (≥50% improvement in MADRS scores from baseline to post-treatment) vs. Non-responders | SVM radial kernel | 79.2% Using baseline EEG data 82.4% Using baseline and week 2 EEG data | Baseline Model Sensitivity—67.3% Specificity—91.0% Baseline and Week 2 Model Sensitivity: 79.2% Specificity: 85.5% |
