On the value of radiomics in addition to clinical measures in emotional conflict fMRI for predicting sertraline response in major depressive disorder Authors: Mingshi Chen, Henk A Marquering, Henricus G Ruhé, Liesbeth Reneman, Matthan WA Caan Published: 2026 ABSTRACT AbstractMost patients with major depressive disorder (MDD) do not respond to first-line treatment, and robust predictors of individual outcomes are lacking. Functional MRI (fMRI) has shown promise for identifying treatment-relevant brain markers, but standard analyses often rely on mean activation within regions of interest (ROIs). Emotional-conflict fMRI elicits spatially heterogeneous, potentially clinically relevant activation, which radiomics can capture by quantifying the texture and distribution of voxel-wise activity. In this preregistered study, we tested whether radiomic features improved the prediction of treatment response compared to mean fMRI ROI activation at baseline (week 0) and early-treatment (week 1) of 254 MDD outpatients participating in the EMBARC RCT (123 sertraline, 131 placebo). Across models, clinical variables provided the strongest predictive signal, and imaging-only models showed limited predictive utility. Placebo but not sertraline response could be predicted at week 1 early-treatment with three combined models incorporating different levels of imaging information, although not at week 0 baseline. In our regression analysis, combined models including mean ROI features yielded modest performance for sertraline (baseline: r = 0.22; early-treatment: r = 0.29; P < 0.05), whereas radiomic models underperformed (baseline: P > 0.05), whereas whole-brain models modestly improved prediction (baseline: r = 0.29; P = 0.0012). Early clinical measures remained the strongest predictors with higher importance than radiomic features. These findings suggest limited added value of emotional conflict fMRI or the addition of radiomics features over clinical data for predicting treatment outcomes in MDD.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56688-y.Keywords: Major depressive disorder, Treatment response, Emotional conflict fMRI, Machine learning, RadiomicsSubject terms: Biomarkers, Diseases, Medical research, Neurology, Neuroscience Introduction Major depressive disorder (MDD) is one of the most prevalent and disabling psychiatric disorders, affecting over 280 million people worldwide1,2. Despite the widespread use of antidepressants, response to first-line selective serotonin reuptake inhibitors (SSRIs) remains highly variable, with around only half of patients responding in a first-step sertraline trial3–6. The inability to predict antidepressant response at an individual level remains a major challenge in psychiatric treatment, leading to prolonged suffering, trial-and-error prescribing, and increased healthcare costs. Neuroimaging offers diverse techniques that provide complementary insights into brain structure, function, and chemistry, several of which have shown promise in predicting antidepressant response7. Structural Magnetic Resonance Imaging (MRI), Diffusion Tensor Imaging (DTI), Positron Emission Tomography (PET), Magnetic Resonance Spectroscopy (MRS), and Electroencephalography (EEG) have all been linked to treatment response. However, each of these techniques comes with limitations in spatial, temporal, or molecular resolution as reviewed previously by us and others7–9. Task-based fMRI may be a particularly promising alternative because it captures brain activity patterns during cognitive or emotional tasks that are directly affected by MDD, thereby potentially providing functional biomarkers that are closely related to symptom improvement10. One such task widely used in this context is the emotional conflict task, which requires participants to resolve conflicting affective stimuli and reliably engages key regions implicated in emotional regulation, including the amygdala, anterior cingulate cortex (ACC), and dorsolateral prefrontal cortex (dlPFC)10. Neural activity in these regions during emotional conflict processing has been linked to antidepressant treatment outcomes. For example, Fonzo et al. found that greater downregulation of conflict-responsive regions such as the dlPFC, dorsal ACC, and insula during emotional conflict regulation predicted better clinical outcomes with sertraline treatment in depressed patients in the Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) Randomized Clinical Trial11. Similarly, Williams et al. reported that sertraline responders tend to exhibit lower baseline amygdala activity in response to emotional faces compared to non-responders12. Additionally, Godlewska et al. found that increased pretreatment pregenual anterior ACC (pgACC) activity during emotional processing predicted better SSRI treatment response13. This is consistent with findings from Godlewska et al., who demonstrated that early reductions in amygdala and ACC responses to fearful faces after one week of escitalopram predicted clinical response at six weeks in depressed patients14. Together, these findings underscore the potential of emotion-related neural markers to inform and personalize antidepressant treatment strategies. The above-mentioned studies either rely on mean brain activation estimates within predefined regions of interest (ROIs), or whole-brain voxel-based analyses. However, task-related neural responses may have complex spatial heterogeneity and vary over subjects, based on differences in brain structure and function. A more comprehensive analysis framework could thus provide deeper insights into individual differences in treatment response. Radiomics, a quantitative imaging analysis technique, extracts high-dimensional features that capture spatial complexity, heterogeneity, and texture patterns within neuroimaging data15. Moving beyond regional mean activation, radiomics quantifies intricate higher-order statistical features within brain regions, offering a more detailed representation of neural activity16. Indeed, radiomics was proven successful in oncology, after which it has been applied to neuropsychiatric diseases like Alzheimer’s disease and schizophrenia15,17,18. However, the potential of radiomics for predicting antidepressant response in MDD has not been studied yet. In this preregistered secondary analysis, we compared conventional ROI mean-activation features and ROI-based radiomics features from emotional conflict fMRI for predicting treatment outcome. Our primary hypothesis was that radiomics features would provide added predictive value relative to mean ROI activation. Because clinical measures are strong predictors of outcome, we additionally evaluated whether imaging features (ROI or radiomics) improve prediction beyond clinical variables and whether performance differs between sertraline and placebo arms as a secondary analysis. Results Demographics and descriptive statistics Our preregistered secondary analysis of the EMBARC study included medication-free adults aged 18–65 years, diagnosed with chronic or recurrent MDD, with onset of their first depressive episode before age 30. Participants were randomized to receive either sertraline (50–200 mg/day) or placebo for an initial 8-week treatment phase. The primary clinical outcome was change in HAM-D score (ΔHAM-D), calculated as the difference between baseline and week 8 scores. Treatment response was defined as a ≥ 50% reduction in HAM-D, and remission as a HAM-D score of ≤ 7 at week 8. To investigate neural predictors of treatment outcome, participants performed an emotional conflict task during fMRI scanning at either baseline (week 0) or early treatment (week 1). Analyses focused on adaptive regulation of emotional conflict, in line with prior research11. After applying inclusion and exclusion criteria, 254 participants were retained for the week 0 analysis (Table 1; CONSORT flow diagram in Supplementary Fig. S1), and 229 participants for the week 1 analysis (Supplementary Table S1; Supplementary Fig. S2). All participants presented with moderate-to-severe depressive symptoms at baseline. The sertraline and placebo arms did not differ on any baseline demographic or clinical characteristics, including symptom severity at either time point. Table 1Baseline sociodemographic and clinical variables in week 0. Statistics reflect comparisons of the sertraline and placebo arms.CharacteristicSertraline (N = 123)Placebo (N = 131) N % N % tMann-Whitney UP value**Sex, Female8972.48363.42.350.13MeanSDMeanSDP value**Age (years)32.0*(26.0,48.0)34.0*(24.0,46.0)7843.00.72Body mass index27.2*(22.7,31.6)27.3*(22.0, 31.8)6602.00.90Years of education16.0*(13.0,16.0)16.0*(14.0,16.0)7547.50.51Pretreatment HAM-D score18.64.718.84.30.280.78Posttreatment HAM-D score11.0*(6.0,15.0)12.0*(5.0,18.0)4803.00.50Change in HAM-D score7.77.26.87.0−0.900.37SHAPS score33.75.232.0*(29.0,36.0)7113.50.11MASQ scoreAnxious Arousal16.0*(13.0,22.0)16.0*(13.0,19.0)7515.50.41Anhedonic Depression45.0*(41.0,47.0)45.0*(41.0,47.3)7688.00.60General Distress32.58.032.08.1−0.410.68N%N% P value**OutcomeRemission3528.53627.50.230.64Response4335.03929.81.620.20*: Median and interquartile range were obtained if not normally distributed.**: For categorical variables, P values were obtained by Pearson Chi-Square; for continuous variables, P values were obtained by Mann-Whitney U or independent two-sided t-tests. Baseline sociodemographic and clinical variables in week 0. Statistics reflect comparisons of the sertraline and placebo arms. *: Median and interquartile range were obtained if not normally distributed. **: For categorical variables, P values were obtained by Pearson Chi-Square; for continuous variables, P values were obtained by Mann-Whitney U or independent two-sided t-tests. We employed both classification and regression analyses. This dual approach enhances clinical interpretability by supporting binary decision-making (distinguishing responders from non-responders) as well as individualized outcome prediction (estimating the magnitude of symptom change), as well as better comparability to the existing literature. Our preregistered analysis plan (10.17605/OSF.IO/7BWXD), was structured into three models to systematically evaluate predictive performance19. In the ‘conventional ROIs’ model (model A), we use the mean BOLD activation of the post-incongruent incongruent trials vs. post-congruent incongruent trials (iI - cI trials) of emotional-conflict fMRI images in key ROIs, i.e., the bilateral amygdala, dorsal ACC (dACC), pgACC, subgenual ACC (sgACC), dlPFC, insula, and hippocampus (fourteen total), along with clinical variables (see Supplementary Table S9)7,20. The ROIs were chosen based on evidence from previous literature7,10. The ‘radiomics ROIs model’ (model B) incorporates 1302 radiomics features in the 14 ROIs of the emotional-conflict fMRI images. In this model, we capture radiomic features like spatial complexity, being first-order statistical measures and texture features, supplemented by the same clinical variables. Finally, model C, the ‘Conventional whole-brain model’, further expands on model A by including mean activation across all 432 brain regions, also in combination with the same set of clinical features. Our sample size was insufficient to calculate a ‘radiomics whole-brain model’ (due to high risk of overfitting using 93 features across 432 brain regions). The mean outcomes of our analysis are shown in Table 2. The mean activation maps for iI - cI contrast were shown in the Supplementary Fig. S3-6. Table 2Emotional conflict fMRI response performance prediction. Sertraline and placebo groups were assessed at week 0 and 1, (A) for mean task activation and (B) radiomics features in selected ROIs, and (C) mean activation in whole-brain. Area Under the Receiver-Operator-Curve (AUROC) and balanced Accuracy (bAcc) are given.EndpointGroupConventional ROIs model (A)Radiomics ROIs model (B)Conventional whole-brain model (C)AUROCbAccP valueAUROCbAccP valueAUROCbAccP valueSertralineWeek 00.480.510.680.400.460.980.510.520.41Week 10.580.540.0770.590.510.0410.540.530.22PlaceboWeek 00.580.590.0590.550.510.180.590.550.045Week 10.730.65< 0.0010.740.66< 0.0010.750.66< 0.001AUROC, area under the receiver operating characteristic curve. bACC, balanced Accuracy. P values were obtained using a one-tailed DeLong test. Emotional conflict fMRI response performance prediction. Sertraline and placebo groups were assessed at week 0 and 1, (A) for mean task activation and (B) radiomics features in selected ROIs, and (C) mean activation in whole-brain. Area Under the Receiver-Operator-Curve (AUROC) and balanced Accuracy (bAcc) are given. AUROC, area under the receiver operating characteristic curve. bACC, balanced Accuracy. P values were obtained using a one-tailed DeLong test. Classification model performance Classification model A demonstrated modest predictive accuracy for treatment response. In the sertraline arm, performance was lower than placebo, with AUROC values increasing from 0.48 at week 0 to 0.58 at week 1 and corresponding balanced accuracies of 0.51 and 0.54. In the placebo arm, AUROC improved from 0.58 at week 0 to 0.73 at week 1, with balanced accuracies of 0.59 and 0.65, respectively (see Table 2). Classification results for remission showed a similar pattern of improvement from week 0 to week 1 (see Supplementary Table S2). In contrast to our hypothesis, incorporating radiomics features in classification model B did not lead to significant performance gains (i.e., week 0 placebo model: AUROC = 0.55, P = 0.51) but rather similar results. Classification model C was broadly comparable to model A. The only statistically significant difference between the three models was found in the placebo week 0 condition, where model A significantly outperformed model B in only remission models (P = 0.013; see Supplementary Table S4) but not response models (P = 0.17; McNemar test, see Supplementary Table S3). Contrary to our hypothesis, the placebo group consistently showed numerically higher AUROC values than the sertraline group across most conditions for both the response and remission models. This suggests a potentially better ability to predict placebo response. However, these differences between the sertraline and placebo groups were not statistically significant (P > 0.05). Feature importance analysis revealed that model predictions were primarily driven by clinical variables, which accounted for nearly all of the top ten predictors. In the classification model A that predict sertraline response using week 1, the most influential features were the MASQ Anhedonic Depression score at week 1, the relative reduction in HAM-D-17 score from week 0 to week 1, and body mass index (BMI); for placebo were HAM-D-17 score in week 1, MASQ General Distress score in week 1 and relative reduction in HAM-D-17 score (see Supplementary Fig. S7 and S8). Features from the model B and the model C are not described, as they either closely resembled those in the model A or demonstrated limited predictive value. Regression model performance Regression model A demonstrated statistically significant, though modest, predictive performance (see Table 3). For the sertraline arm, Pearson correlations were r = 0.22 at week 0 and r = 0.29 at week 1 (both P < 0.05), corresponding to R² values of 1.0% and 7.4%, with RMSEs of 7.05 and 7.11, respectively (see Table 3). In the placebo arm, the regression model A at week 1 was statistically significant (P = 0.018), while the week 0 model was not. Interestingly, predictive performance for both treatment arms was slightly better at week 1 than at week 0. Table 3Response Outcome prediction performance for the regression models on validation data.EndpointGroupConventional ROIs model (A)Radiomics ROIs model (B)Conventional whole-brain model (C)Pearson CorrelationP valueR2RMSEPearson CorrelationP valueR2RMSEPearson CorrelationP valueR2RMSESertralineWeek 00.220.0141.0%7.050.120.190.4%7.070.290.00127.3%6.82Week 10.290.00167.4%7.110.230.0113.2%7.270.220.0171.0%7.35PlaceboWeek 00.0740.40< 0%7.24−0.0240.79< 0%7.40−0.220.010< 0%7.31Week 10.220.0182.3%6.910.32< 0.0018.6%6.680.260.00525.6%6.79RMSE, root-mean-square error. P-values were obtained by Pearson Correlation. Response Outcome prediction performance for the regression models on validation data. RMSE, root-mean-square error. P-values were obtained by Pearson Correlation. In the regression model B, performance declined notably in the sertraline week 0 model, which was no longer statistically significant, while other subgroups remained comparable to model A. The regression model C outperformed the model A only in the sertraline week 0 model and placebo week 1 model, achieving r = 0.29 (P = 0.0012) and r = 0.26 (P = 0.0052), but overall, no significant differences were observed among the three models (see Supplementary Table S5). Feature importance analysis emphasized, similarly to classification, the dominant role of clinical variables in predicting treatment outcomes. In the week 1 sertraline regression model A, 60% of the top 20 features were clinical, with the three most influential being the relative reduction in HAM-D from week 0 to week 1, the HAM-D score at week 1, and body mass index (BMI) (see Fig. 1A). Here, of the top 10 features, three were imaging-based. In contrast, the placebo model showed an even stronger reliance on clinical data, with 13 clinical features in the top 20 and all top 10 predictors being clinical (see Fig. 1B). SHAP summary plot for the conventional ROIs regression model A. 1 A: Sertraline week1, training set. 1B: Placebo week 1, training set. Dot color indicates feature value (red = high, blue = low), and horizontal position shows impact on the model’s prediction (left = negative effect, right = positive effect).GD: General Distress; AA: Anxious Arousal; AD: Anhedonic Depression. Ablation experiments In order to evaluate the relative contribution of different feature modalities, we conducted ablation experiments by exclusively selecting either clinical or imaging features from model A. Here, model A was selected given its comparable or higher performance compared to model B and model C in most analyses, with the lowest number of features. In the classification models for treatment response, models using only clinical features performed similarly to the full combined models (See Supplementary Table S6). The AUROC scores for the clinical-only models were 0.43 for sertraline week 0, 0.53 for placebo week 0, 0.59 for sertraline week 1, and 0.75 for placebo week 1. In contrast, models using only mean activation ROIs imaging features performed no better than chance. A similar pattern was observed in both the remission classification and regression models (See Supplementary Table S7, S8), further highlighting the dominant predictive value of clinical features over imaging features (See scatter plots in Supplementary Fig. S9, S10). We also repeated the imaging-only ablation while skipping the ComBat harmonization step. Imaging-only performance remained at chance level across tasks (see Supplementary Tables S6–S8). Discussion We explored the added value of radiomics features from emotional conflict fMRI for predicting treatment response in MDD, and investigated the utility of early-treatment data for improving predictive accuracy. Contrary to our hypothesis, the radiomics ROIs model (B) and the conventional whole-brain model (C) did not outperform the conventional ROIs model (A), which relied on average activation in a pre-defined subset of features. Notably, we were unable to replicate findings from a previous study that used the same (EMBARC) data11, despite applying only slightly different improvements in preprocessing steps. Also unexpectedly, classification models for both response and remission showed no superiority for the sertraline group compared to placebo. In fact, AUROC values for placebo were consistently higher than those for sertraline at both baseline (week 0) and early treatment (week 1). Finally, while imaging features provided some predictive value, clinical variables emerged as the most influential predictors across models. While we followed previous work in most aspects11, we employed more standardized and generalizable methods to date. For instance, to improve model stability and generalizability, we adopted open software of fMRIprep for preprocessing, ComBat for site-effect correction, nested cross-validation, and XGBoost for model training. Indeed, our radiomics analysis achieved an above-average Radiomics Quality Score (RQS), reflecting strong methodological rigor and alignment with recommended best practices in radiomics research15. While Fonzo et al. reported a Pearson correlation of r = 0.49 for predicting sertraline response at week 0 (emotional conflict, iI - cI contrast), our highest correlation at week 0 was only a modest r = 0.29 (conventional whole-brain model C). These findings indicate that predictive modeling using fMRI data can be highly sensitive to even minor differences in analytic pipelines, underscoring the importance of methodological transparency and standardization21,22. Our modest performance in predicting treatment outcome is corroborated by Nguyen et al., whose methodology was more comparable, albeit using a different reward-processing fMRI task23. Their classification models achieved AUROC scores of 0.60 for sertraline remission, 0.62 for sertraline response, 0.65 for placebo remission, and 0.67 for placebo response23. These results are generally in line with the performance of our models using the emotional-conflict task. In contrast, their regression models showed substantially higher R² values—48% for sertraline and 28% for placebo, which may reflect differences in task design or outcome definitions rather than modeling strategy alone23. These findings suggest that model performance is strongly influenced by the interaction between the fMRI task and the clinical endpoint, with additional methodological factors possibly contributing to the observed differences. We prioritized emotional conflict fMRI because it has the clearest EMBARC-specific prior link to antidepressant outcomes11, making it the most directly justified paradigm in this dataset for testing the added value of radiomics, despite evidence that reward tasks can show higher performance23.All in all, our results raise concerns about the predictive value of the emotional conflict task in the context used here. For instance, Alders et al. reported no significant differences in emotional conflict-related activation between MDD patients and healthy controls after eight weeks of escitalopram treatment in the CAN-BIND dataset24. Similarly, Fisher et al. reported that brain responses to emotional faces during an fMRI task neither distinguished MDD patients from controls nor predicted treatment outcomes (AUC-ROC < 0.6) following eight weeks of antidepressant therapy in the NeuroPharm dataset25. Indeed, a meta-analysis encompassing 90 experiments with over 1,000 participants reported poor overall reliability of task-based fMRI, with an average ICC of 0.39726. Such findings cast doubt on the dependability of task-based fMRI studies, including those involving emotional conflict paradigms. The timing of evaluation plays a critical role in our model performance, with early-treatment data (week 1) generally offering stronger predictive power than pre-treatment data (week 0). This improvement may be partly explained by differences in information content: unlike week 0 models, week 1 models incorporated additional early-treatment information, including week 1 clinical measures and week 0 to week 1 change scores (e.g., relative HAM-D reduction). Consistent with this, these early symptoms and change measures were frequently among the most influential predictors, suggesting that the observed performance gains likely reflect the added longitudinal/early-response signal rather than a strictly time-matched comparison. The ability to predict treatment response early in the course of therapy holds significant clinical value. Although the full therapeutic effects of antidepressants typically emerge after 4–6 weeks of treatment27, and the EMBARC study evaluated outcomes over an 8-week period28, our results suggest29 that meaningful clinical signals may already be detectable within the first week. Our findings align with prior literature showing that early improvement—within the first 2 weeks—can predict later treatment outcomes with high sensitivity in patients with MDD29,30. These early indicators could support more personalized and effective treatment decisions. Nonetheless, further evidence is needed to confirm the utility of such early predictions in real-world clinical settings. In most cases, our simpler model A (mean activation within predefined ROIs) either outperformed or matched the performance of more complex models B and C, both in classification and regression tasks. While radiomic methods offer detailed and high-dimensional features, they did not improve predictive performance in this context. Importantly, the emotional conflict task, across all models, failed to yield robust imaging-based predictors of treatment outcome. In the ablation study, imaging-only models performed at chance level, while clinical-only models achieved similar accuracy to the combined models. This pattern suggests that the apparent disadvantage of the radiomics model is unlikely to reflect meaningful imaging signal, but may instead stem from methodological factors inherent to high-dimensional modeling—most notably the large initial feature space and multicollinearity relative to the available sample size, which can increase estimation variance and impede generalization even after feature reduction. These findings suggest that, for this specific task and outcome, emotional conflict fMRI features do not provide meaningful predictive value beyond what is captured by clinical measures, consistent with previous findings that predictive performance improves primarily when clinical features are included20,23. Indeed, clinical measures, particularly the HAM-D score, relative HAM-D reduction, MASQ score and BMI, consistently emerged as the strongest and most reliable predictors of treatment response at week 1 in both regression and classification models across sertraline and placebo groups. These clinical measures are not only more predictive but also practical, cost-effective, and easily implementable in clinical practice. Of note, we previously demonstrated that quantitative MRI measures assessing resting-state hemodynamics (e.g., brain perfusion) can outperform clinical assessments at baseline20. Taken together, these findings highlight that while advanced quantitative imaging techniques, including resting-state perfusion and other hemodynamic MRI measures, may enhance predictive accuracy at treatment initiation, straightforward clinical measures continue to provide robust, clinically meaningful guidance for personalizing MDD treatment. The placebo group exhibited a relatively high response rate of 29.8% at week 8, which is somewhat comparable to the sertraline group’s response rate of 35.0%. This finding is consistent with the literature, which reports pooled placebo response rates in depression ranging from 30% to 50%31,32. Previous studies have shown that placebo effects may involve neurobiological mechanisms, such as the conditioned release of oxytocin, although these effects tend to diminish over time33. Overall, these findings highlight the need to consider placebo effects when interpreting treatment outcomes. A further implication of the comparable week 8 outcomes between sertraline and placebo is that detecting strictly drug-specific biomarkers may be difficult in this dataset. The high placebo response suggests that much of the predictable variance may reflect general improvement processes rather than sertraline-specific mechanisms. This interpretation is consistent with our findings that clinical variables were the strongest predictors across models, while emotional conflict fMRI features provided limited incremental value. Therefore, the primary clinical message is not the identification of a sertraline-specific imaging biomarker, but rather that routinely available clinical measures offer more robust and practical prediction of outcome than the imaging features tested here. Limitations of the study include the site and scanner heterogeneity. Although ComBat was used to mitigate scanner-related biases, it may obscure the true biological signal and cannot fully address nonlinear effects. In a post-hoc ablation study omitting ComBat, outcomes did not improve (see Supplementary Table S6-8). Our models were developed for the sertraline and placebo arms only, limiting generalizability to other treatments. Model B began with 1,302 radiomic features, which poses a high-dimensional, moderate sample size setting and increases the risk of unstable feature selection. Even though we reduced the feature set using filtering and recursive feature elimination (RFE), this dimensionality reduction may have removed subtle distributed signal and may have contributed to the limited performance of Model B. We also lacked external validation in an independent cohort. Therefore, conclusions, especially the null added value of radiomics, should be interpreted as specific to EMBARC and our preprocessing and modeling framework. The preregistration included additional exploratory analyses (alternative algorithms and normative modeling), which are not reported here because the primary models showed limited predictive utility. Lastly, while predictive, the models offer limited mechanistic insight; understanding the neurobiological basis will require complementary analyses, such as mediation models or integration with molecular data. In conclusion, our findings suggest that, in contrast to our initial hypothesis, neither emotional conflict fMRI nor the additive radiomics features derived from this task added predictive value for treatment response in MDD. The lack of replication of prior EMBARC-based results underscores the methodological sensitivity of machine learning in neuroimaging, with possible differences introduced by variation in preprocessing pipelines, feature selection strategies, and analytic frameworks. While imaging contributed modestly, clinical variables consistently outperformed neuroimaging features, emphasizing their central role in outcome prediction. These results highlight the need for careful methodological standardization of post-acquisition workflows to improve reproducibility in neuroimaging-based prediction studies. Methods Participants and study design We conducted a preregistered secondary analysis19 of data from the EMBARC study, a multi-site, double-blind, placebo-controlled randomized clinical trial (ClinicalTrials.gov identifier, NCT01407094) designed to investigate biomarkers predictive of antidepressant response in patients with major depressive disorder (MDD)28. The original EMBARC trial enrolled medication-free adults aged 18–65 years who met diagnostic criteria for chronic or recurrent MDD, experienced their first depressive episode before age 30, and had a Quick Inventory of Depressive Symptomatology score of at least 14 at both screening and randomization. Key exclusion criteria included non-response to prior antidepressant treatment during the current depressive episode, current pregnancy, lifetime history of psychosis or bipolar disorder, recent substance abuse or dependence, unstable psychiatric or general medical conditions, clinically significant laboratory abnormalities, and contraindications for MRI scanning (e.g., epilepsy, metallic implants). A detailed description of study methods and participant characteristics is available in the original EMBARC design publication28. Written informed consent was obtained from all participants under protocols approved by the Institutional Review Board (IRB) at each of the four clinical sites: University of Texas Southwestern Medical Center, Massachusetts General Hospital, Columbia University, and the University of Michigan. All research was conducted in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. Further exclusions were applied for missing behavioral or emotional conflict fMRI data, poor behavioral performance (accuracy below 80%), and excessive motion during fMRI scans (mean framewise displacement (FD) above 0.4 mm)20. A CONSORT Flow Diagram is presented in Supplement Fig. S1 for week 0 and Fig. S2 for week 1. Participants were randomized to receive sertraline (50–200 mg/day) or placebo for an initial 8-week treatment phase. Clinical response was assessed using the HAM-D at baseline and at the end of each 8-week treatment phase. Primary clinical outcomes included the change in HAM-D scores (ΔHAM-D), calculated by subtracting week 8 scores from baseline scores (ΔHAM-D = week 0 – week 8). Treatment response was defined as a ≥ 50% reduction in HAM-D scores, and remission was operationalized as achieving a HAM-D score of ≤ 7 by the end of 8 weeks of treatment. Emotional conflict task and fMRI preprocessing Participants performed an emotional conflict task during fMRI scanning at baseline (week 0) or early treatment (week 1), involving identification of emotional facial expressions (happy or fearful) overlaid with either congruent or incongruent emotional words (“HAPPY” or “FEAR”). Trials consisted of congruent (matching word and face) and incongruent conditions (mismatched word and face), capturing emotional conflict processing and implicit emotional conflict regulation. Based on prior literature, we analyzed iI - cI trials to investigate adaptive regulation of emotional conflict11. Data were acquired at four different sites on 3 T MRI scanners from different vendors, using a gradient echo echo-planar imaging fMRI protocol of 13 min and 14 s. Detailed neuroimaging acquisition parameters are described in Supplementary Table S11. fMRI data were preprocessed using fMRIPrep 24.1.134, including head motion correction, slice-timing correction, spatial normalization to the Montreal Neurological Institute (MNI152) template, spatial smoothing (6 mm Full Width at Half Maximum (FWHM)), and calculating the confounds and removal of motion-related independent components with ICA-AROMA (see Supplementary Fig. S11). Individual-level activation maps were derived via general linear modeling (GLM) in FSL FEAT 6.0.335,36, contrasting incongruent versus congruent trials. Nuisance signals from segmented white matter and cerebrospinal fluid (CSF) were regressed out. Task events were modeled as zero-duration regressors marking stimulus onset, categorized by congruency (Incongruent, Congruent) and preceding trial type (Post-incongruent, Post-congruent), capturing conflict response and regulation effects separately. Additional nuisance regressors accounted for error and post-error trials, excessive head motion, and six motion parameters. Individual activation maps for the iI - cI contrast were computed. Notably, five subjects from Columbia University had emotional conflict fMRI data with a total of 400 volumes, in contrast to the standard 397 volumes observed in the rest of the sample. To account for this difference, a modified FSL FEAT model was applied specifically to these five participants. After preprocessing, each subject’s data were normalized using the mean functional MRI image generated at their respective site. Feature extraction and missing data imputation Radiomic features were extracted from predefined ROIs using PyRadiomics (version 3.0.1)37. ROI masks were created using the Schaefer 400 cortical and Melbourne 32 subcortical atlases. Based on prior researches7,10, predefined ROIs included bilateral amygdala, dACC, pgACC, sgACC, dlPFC, insula, and hippocampus. Prior to feature extraction, activation maps were normalized by mean intensity. Extracted features included First Order Statistics (19 features), Gray Level Co-occurrence Matrix (24 features), Gray Level Run Length Matrix (16 features), Gray Level Size Zone Matrix (16 features), Neighboring Gray Tone Difference Matrix (5 features), and Gray Level Dependence Matrix (14 features), totaling 93 features per ROI (see Supplementary Table S10 for more details). In the conventional ROIs model, we used 14 ROIs with corresponding mean activation features. Building on this, the radiomics ROIs model extracted a total of 1,302 radiomic features from the same 14 ROIs. For the conventional whole-brain model, we combined the Schaefer 400-parcel cortical atlas with the Melbourne 32-region subcortical atlas, yielding 432 mean activation features in total. We included clinical and sociodemographic variables because they are clinically accessible and have demonstrated substantial predictive value in EMBARC, including in recent multimodal work where clinical measures (especially early symptom change) were among the strongest predictors and improved performance beyond imaging alone20. Baseline (Week 0) clinical variables included Body Mass Index (BMI), Mood and Anxiety Symptom Questionnaire (MASQ) scores, Snaith-Hamilton Pleasure Scale (SHAPS) scores, baseline HAM-D score, age, gender, employment status, and chronicity. Additional Week 1 variables included HAM-D score, relative HAM-D score reduction from baseline (Week 1/Week 0 ratio), and SHAPS and MASQ scores at Week 1 (See Supplementary Table S9). Missing data in HAM-D outcomes at week 8 were imputed using iterative multivariate conditional modeling from sklearn 1.6.1 IterativeImputer based on baseline HAM-D, week 1 HAM-D, week 2 HAM-D, week 3 HAM-D, week 4 HAM-D, week 6 HAM-D, baseline Quick Inventory of Depressive Symptoms (QIDS) total score, baseline MASQ scores for Anxious Arousal, Anhedonic Depression, and General Distress, SHAPS total score, age, years of education, gender, and Wechsler Abbreviated Scale of Intelligence (WASI) t-scores for Vocabulary and Matrix Reasoning. Incomplete or missing data in HAM-D outcomes at week 1, MASQ, and SHAPS scores were imputed using iterative multivariate conditional modeling based on week 1 to week 6 HAM-D scores, MASQ week 0 and week 1, and SHAPS scores week 0 and week 1, respectively. BMI was imputed using the median value across the cohort, and employment status was imputed using the most frequent category. Model training and evaluation To comprehensively evaluate model performance, we employed both classification and regression analyses. This dual approach enhances clinical interpretability by supporting binary decision-making (e.g., distinguishing responders from non-responders) as well as individualized outcome prediction (e.g., estimating the magnitude of symptom change). While some studies focus exclusively on classification metrics and others on regression outcomes, incorporating both improves the comparability of our findings within the broader literature. Predictive modeling was conducted using eXtreme Gradient Boosting (XGBoost, version 3.0.0). Prior to modeling, features were first imputed as described above, then harmonized to correct for site effects using the neuroHarmonize method (i.e., ComBat, version 2.2.0). Although XGBoost can natively handle missing values, we imputed missing features prior to harmonization because ComBat (neuroHarmonize) requires a complete feature matrix and does not support missing entries. The imputed matrix was then harmonized and used for downstream modeling. ComBat harmonization was applied only to imaging-derived features to mitigate site effects; clinical variables were not harmonized and were included unchanged in the modeling pipeline. Subsequently, features were scaled using the RobustScaler from scikit-learn (version 1.6.1), which adjusts values based on the interquartile range (25th to 75th percentiles). Feature selection was implemented using a custom transformer that applies three sequential steps. First, features containing any missing values were excluded. Second, low-variance features were removed using median absolute deviation (MAD), with a threshold of MAD ≤ 0.001. Third, highly correlated features were removed using a pairwise Pearson correlation threshold of |r| ≥ 0.8. For each highly correlated pair, we removed the feature with the higher mean absolute correlation across all remaining features. After these filtering steps, we applied RFE with an XGBoost estimator, removing one feature per iteration until 10 features remained. This target dimensionality was chosen a priori to keep model complexity aligned with the available sample size (≈) and follows the recommended optimized feature number reported by Hua et al.38. For classification tasks, model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUROC), while regression performance was assessed with root-mean-square error (RMSE). Hyperparameter tuning was performed using BayesSearchCV from scikit-optimize (version 0.9.0) across 50 iterations. To address class imbalance, SMOTE (Synthetic Minority Over-sampling Technique) was applied only for classification tasks and was not used for regression analyses. To ensure model robustness and generalizability, we employed a nested cross-validation strategy using KFold with 10-fold inner and 10-fold outer loops, implemented in scikit-learn 1.6.1 (see Fig. 2 for a schematic of the analysis pipeline). Feature importance was assessed using SHapley Additive exPlanations (SHAP, version 0.46.0). Flow chart of machine learning pipeline.* ComBat harmonization was applied only to imaging-derived features to mitigate site effects; clinical variables were not harmonized and were included unchanged in the modeling pipeline. ** SMOTE (Synthetic Minority Over-sampling Technique) was used only for classification tasks to address class imbalance, and was not applied to regression analyses. For classification tasks, performance was assessed using the overall AUROC and balanced accuracy (bAcc). Statistical significance was assessed using a one-tailed DeLong test on the AUROC, as AUROC provides a more informative measure of classification performance by capturing discrimination across all possible decision thresholds. A one-tailed test was used because the hypothesis of interest was directional, namely whether the model performed better than chance rather than simply differently from chance. Accordingly, we tested the null hypothesis that AUROC ≤ 0.5 against the alternative hypothesis that AUROC > 0.5. For regression models predicting change in HAM-D scores (ΔHAM-D), model performance was evaluated using R-squared and RMSE, with statistical significance assessed via Pearson correlation. Differences among models were analyzed using Friedman’s test and pairwise Wilcoxon tests for regression models, and McNemar’s test for classification models. A significance threshold of p < 0.05 was applied throughout. Preregistration note (exploratory analyses) The preregistration described several exploratory analyses, including comparing additional algorithms (e.g., linear/regularized regression and BART), normative modeling using the placebo arm to estimate drug-specific improvement, and additional subgroup analyses of non-responders. These exploratory analyses are not reported in the present manuscript because the primary preregistered models showed limited predictive utility and did not support robust imaging-based prediction. We therefore prioritized reporting the main preregistered model comparisons and interpretation of these results. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements This research has received funding support from the China Scholar­ship Council (CSC), No.202206380049.The EMBARC study was supported by NIMH under grants U01MH092221 and U01MH092250. The EMBARC study was also supported in part by Valeant Pharmaceuticals, which donated extended-release bupropion, used in the second phase of the EMBARC study (not reported here). Author contributions Mingshi Chen: Writing – original draft, Visualization, Validation, Investigation, Funding acquisition, Formal analysis. Henk A. Marquering: Writing – review & editing, Supervision. Henricus G. Ruhé: Writing – review & editing, Supervision. Liesbeth Reneman: Funding acquisition, Writing – review & editing, Supervision, Resources, Project administration, Data curation, Conceptualization. Matthan W.A. Caan: Funding acquisition, Writing – review & editing, Supervision, Software. Data availability The authors confirm that the data supporting the findings of this study are available within thearticle and its supplementary material. The documented code base is available on GitHub(https://github.com/Mancy-Chen/EMBARC-fMRI). Also, the trained models and selected features have been uploaded to the OSF platform (https://doi.org/10.5281/zenodo.19069258).Raw MRI data to support the findings of this study are available through the NIMH Data Archive(https://nda.nih.gov/edit\_collection.html? id=2199). Declarations Competing interests M.W.A. Caan is a shareholder of Nico-lab International Ltd. H.A. Marquering is a shareholder of Nico-lab International Ltd., TrianecT B.V., and inSteps B.V. All other authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Liesbeth Reneman and Matthan W.A. Caan these authors contribution equally this work. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The authors confirm that the data supporting the findings of this study are available within thearticle and its supplementary material. The documented code base is available on GitHub(https://github.com/Mancy-Chen/EMBARC-fMRI). Also, the trained models and selected features have been uploaded to the OSF platform (https://doi.org/10.5281/zenodo.19069258).Raw MRI data to support the findings of this study are available through the NIMH Data Archive(https://nda.nih.gov/edit\_collection.html? id=2199).