A new task-fMRI study published in Frontiers in Human Neuroscience takes exactly that approach, and its central finding is striking in its simplicity. When researchers built a leakage-controlled multimodal machine learning framework to classify adolescents with BPD against healthy controls, the single most informative signal came not from resting-state network architecture and not from clinical questionnaires, but from the brain’s moment-to-moment responses during an emotional Go/No-Go task. Task-evoked features derived from a general linear model analysis achieved an area under the curve of 0.881 ± 0.055 as a single modality, while the best functional connectivity model reached only 0.664 ± 0.053 and the best clinical-only model 0.776 ± 0.067. The full multimodal model, combining all three feature domains, achieved an AUC of 0.883 ± 0.048 with an accuracy of 0.816 ± 0.068, only marginally above the task-evoked result.

The study, led by Mingzhao Ma of The Second Xiangya Hospital of Central South University together with colleagues across Chinese and Italian institutions, recruited 258 adolescents aged 12 to 17 years from outpatient clinics at the Mental Health Center of Xiangya Hospital, including 129 individuals with BPD and 129 healthy controls. After excluding seven BPD participants with missing clinical data, the final analytic sample comprised 251 adolescents. BPD diagnosis required fulfillment of at least five of nine DSM-IV criteria, symptom stability of at least two years not fully explained by Axis I disorders, and a score above 66 on the Borderline Personality Feature Scale for Children. Controls had no psychiatric history and an IQ above 80. Exclusion criteria for both groups covered schizophrenia spectrum, bipolar spectrum, post-traumatic stress disorder, neurodevelopmental disorders, major depressive disorder, substance dependence, neurological disorders, and IQ at or below 80.

During scanning on a Siemens MAGNETOM Prisma 3T scanner, participants performed an emotional Go/No-Go task adapted from Wessa and colleagues, requiring them to press a button when happy, sad, or neutral target faces appeared and to withhold responses when distractor faces appeared. Each facial stimulus was shown for 500 milliseconds with a 1,000-millisecond inter-stimulus interval, across six emotional blocks and two gender-control blocks. The primary contrast of interest was happy No-Go versus neutral No-Go, a comparison designed to isolate the unique interference of positive emotional arousal on response inhibition. Happy faces were chosen deliberately because they represent real-life social and reward cues that are particularly salient for adolescents, allowing the team to probe whether adolescents with BPD were more prone to losing control or acting impulsively when confronted with positive emotional triggers.

The machine learning pipeline was engineered with unusual care. Three feature domains were constructed: functional connectivity matrices vectorized from their upper triangular elements, task-evoked GLM-derived temporal features capturing ten statistics of regional BOLD time courses including mean, standard deviation, percentiles, and first-difference dynamics, and 38 clinical variables spanning demographic, symptom-level, cognitive, and task-behavioral measures. Variables that could directly encode diagnostic grouping, such as disease duration and age at onset, were excluded to prevent shortcut learning. Seven feature settings, from single-modality through full multimodal combinations, were evaluated across seven conventional classifiers including logistic regression, support vector machine, k-nearest neighbors, Gaussian naive Bayes, multilayer perceptron, random forest, and XGBoost or gradient boosting, all under the same five-fold cross-validation protocol.

What sets this study apart methodologically is its strict leakage control. In neuroimaging classification, information leakage occurs when feature selection, normalization, dimensionality reduction, or region definition is performed using the full dataset before cross-validation, inflating performance estimates. Here, task-evoked GLM regions were identified separately within each cross-validation fold using training subjects only, with significant clusters thresholded at alpha 0.05 and a cluster threshold of 10 voxels. Feature standardization, ANOVA-based feature selection retaining up to 256 connectivity features, PCA retaining up to 32 components, and classifier fitting were all performed strictly within training folds and then applied to held-out test participants without using test labels or distributional information. Hyperparameters were prespecified rather than tuned on test folds, reducing optimization bias at the cost of potentially underestimating tuning-sensitive algorithms.

SHAP analysis, computed independently within each fold using predictions for held-out test participants only, confirmed the predominance of task-evoked features. GLM-derived features accounted for 83.1 percent of total model attribution in the full multimodal random forest model, compared with 9.5 percent for functional connectivity and 7.5 percent for clinical variables. At the neural level, the most influential regions were distributed across visual and perceptual areas including the calcarine cortex, cuneus, and occipital cortex, temporoparietal default-mode hubs including the angular and supramarginal gyri, frontal control regions including the inferior frontal gyrus, and subcortical structures including the putamen. This distributed pattern suggests a hierarchical organization spanning perceptual input, contextual and self-referential processing, and top-down inhibitory control systems.

The authors interpret these findings through the lens of the hierarchical affective predictive imbalance account, a predictive-processing framework proposed by Herzog and colleagues suggesting that imbalanced hierarchical priors may contribute to the affective instability and impulsivity characteristic of BPD. Within this framework, functional connectivity captures the baseline organization of predictive systems, whereas task-evoked responses reflect their functional deployment during active predictive regulation. The dominance of task-evoked features suggests that disorder-relevant information in adolescent BPD becomes most evident when the brain is required to dynamically regulate predictions and actions under affective interference, rather than in intrinsic network organization alone. The prominence of perceptual and temporoparietal regions points to altered processing and contextual integration of emotionally salient stimuli, while frontal and subcortical involvement reflects disruptions in inhibitory control and response selection.

Yet the authors are careful about what their data can and cannot establish. Because the study does not directly estimate prediction errors, precision weighting, or hierarchical belief updating, it is not a mechanistic test of predictive processing. The same regional pattern may equally reflect visual attention, salience processing, and executive inhibitory control, conventional accounts that provide plausible alternative explanations. SHAP-derived regional importance should be read as predictive attribution rather than causal evidence, since model attribution identifies features that contribute to classification decisions without establishing whether those regions are mechanistically responsible for symptoms. Several SHAP-relevant regions overlap with systems undergoing protracted adolescent maturation, including frontostriatal, salience, and temporoparietal networks, and future work integrating task-fMRI biomarkers with transcriptomic or molecular datasets may clarify the biological pathways underlying these disease-relevant neural states.

Behavioral performance analyses added nuance to the neural findings. After Benjamini-Hochberg correction, the BPD group showed lower neutral Go omission-error percentages and lower neutral No-Go commission-error percentages than controls, with Hedges g effect sizes of −0.460 and −0.311 respectively, while happy-condition measures and reaction times did not differ significantly. The BPD group therefore did not show a generalized behavioral deficit across task conditions. An age- and sex-matched sensitivity analysis, forming 89 exact pairs with identical mean age and female proportion, yielded held-out AUCs of 0.866 for the GLM random forest model and 0.869 for the full multimodal model, reducing but not eliminating concern that demographic imbalance explains classification. The parent clinical cohort included psychotropic medication exposure in 22.5 percent of BPD participants, and residual medication effects on task-evoked signals cannot be excluded.

The translational implications are considerable. Task-based neural responses constitute a particularly informative and interpretable biomarker for individualized BPD identification, and the strong performance of GLM-derived features suggests that task-evoked dynamics already capture most of the disease-relevant signal in this cohort. Different classifiers emphasized different diagnostic trade-offs: random forest achieved the highest AUC and specificity, favoring conservative identification with fewer false positives, while k-nearest neighbors achieved the highest accuracy, F1 score, and recall, offering a more balanced sensitivity-specificity profile. In screening contexts, higher sensitivity may be prioritized to minimize missed cases, whereas confirmatory or decision-support settings may favor the higher specificity and precision of ensemble models. The authors emphasize that independent external validation remains necessary to establish cross-site translational robustness, and that definitive algorithm ranking would require fully nested hyperparameter optimization in an independent or larger cohort. Still, the convergence of leakage-controlled classification, cross-validated SHAP attribution, and a theoretically coherent hierarchical neural pattern marks a meaningful step toward capturing disorder-relevant brain states at the individual level in adolescent psychiatry.

Subject of Research: Neuroscience

Article Title: Hierarchical affective predictive imbalance in adolescent borderline personality disorder: task-evoked fMRI evidence from multimodal machine learning

Article References: Ma, M., Wang, X., Tan, Z., Liu, J., Xiao, Q., Hu, T., Grecucci, A., & Yi, X. (2026). Hierarchical affective predictive imbalance in adolescent borderline personality disorder: task-evoked fMRI evidence from multimodal machine learning. Frontiers in Human Neuroscience, 20, Article 1875156. https://doi.org/10.3389/fnhum.2026.1875156

Image Credits: AI Generated

DOI: 10.3389/fnhum.2026.1875156

Keywords: adolescent BPD, task-fMRI, machine learning, predictive processing, neuroimaging biomarkers, emotional Go/No-Go, Hierarchical, affective, predictive, imbalance, adolescent, borderline