Hierarchical affective predictive imbalance in adolescent borderline personality disorder: task-evoked fMRI evidence from multimodal machine learning
6 affiliations
- 1Department of Radiology, The Second Xiangya Hospital of Central South University, Central South University , , ,
- 2Department of Nuclear Medicine, Chongqing University Three Gorges Hospital, Chongqing University , ,
- 3Department of Neurosurgery, Xiangya Hospital, Central South University , , ,
- 4Mental Health Center of Xiangya Hospital, Central South University , , ,
- 5School of Computer Science and Engineering, Central South University , , ,
- 6Department of Education, Psychology and Communication Sciences, University of Bari , ,
Abstract
Background
Predictive-processing accounts of BPD, including the hierarchical-prior-imbalance account proposed by Herzog, offer a framework for interpreting affective instability and impulsivity. The present task-fMRI study evaluates whether discriminative task-evoked patterns are consistent with this framework; because it does not directly estimate prediction errors, precision weighting, or hierarchical belief updating, it is not a mechanistic test of predictive processing.Methods
To evaluate this question, this study developed a leakage-controlled multimodal machine learning framework for individualized classification of adolescents with BPD and healthy controls based on emotional Go/No-Go task-fMRI activity. Three feature domains were constructed: functional connectivity (FC), task-evoked general linear model (GLM)-derived temporal features, and clinical variables. Seven feature settings, including single-modality, dual-modality, and full multimodal combinations, were evaluated using multiple conventional machine learning classifiers, including logistic regression, support vector machine, k-nearest neighbors, Gaussian naive Bayes, multilayer perceptron, random forest, and XGBoost or gradient boosting.Results
The results showed that task-evoked GLM-derived features provided the strongest single-modality discriminative information (AUC of 0.881 ± 0.055). SHAP analysis confirmed the predominance of GLM-derived features, accounting for 83.1% of total model attribution, compared to 9.5 and 7.5% for functional connectivity and clinical variables, respectively. At the neural level, discriminative features were distributed across visual/perceptual regions (calcarine, cuneus, occipital cortex), temporoparietal/default-mode hubs (angular and supramarginal gyri), frontal control regions (inferior frontal gyrus), and subcortical structures (putamen). This distributed pattern aligns with a hierarchical organization spanning perceptual input, contextual/self-referential processing, and top-down inhibitory control systems, consistent with the proposed model of hierarchical affective predictive imbalance.Conclusion
Task-evoked GLM-derived features were the primary source of discriminative information for adolescent BPD, with functional connectivity and clinical variables providing modest additional value. The resulting classification pattern is consistent with hierarchical predictive-processing accounts but does not establish a specific predictive-processing mechanism. The same regional pattern may also reflect visual attention, salience processing, and executive inhibitory control.
Related
- Normative and Maladaptive Personality Traits in Borderline Personality Disorder: A Comparative and Hierarchical Analysis of HEXACO and PID-5
- Canine-Assisted Interventions and Symptoms Associated with Borderline Personality Disorder: A Scoping Review
- Machine Learning in the Differential Diagnosis of Bipolar and Borderline Personality Disorders: A Scoping Review
- Borderline personality disorder, a diagnosis that cannot be challenged: If rejecting the diagnosis is a symptom, can the diagnosis ever be wrong? A response to Chazhoor and Younger (2026)
- Association of Borderline Personality Traits and Impulsivity with Problem Gambling Severity in Gambling Disorder
BibTeX
@article{ma2026hierarchical,
title = {{Hierarchical affective predictive imbalance in adolescent borderline personality disorder: task-evoked fMRI evidence from multimodal machine learning}},
author = {Mingzhao Ma and Xueying Wang and Zeming Tan and Jun Liu and Qian Xiao and Ting Hu and Alessandro Grecucci and Xiaoping Yi},
journal = {Frontiers in human neuroscience},
year = {2026},
month = oct,
doi = {10.3389/fnhum.2026.1875156},
url = {https://europepmc.org/article/PMC/PMC13716863},
note = {CC BY},
}
RIS
TY - JOUR
TI - Hierarchical affective predictive imbalance in adolescent borderline personality disorder: task-evoked fMRI evidence from multimodal machine learning
AU - Mingzhao Ma
AU - Xueying Wang
AU - Zeming Tan
AU - Jun Liu
AU - Qian Xiao
AU - Ting Hu
AU - Alessandro Grecucci
AU - Xiaoping Yi
PY - 2026
DA - 2026/10/15
JO - Frontiers in human neuroscience
DO - 10.3389/fnhum.2026.1875156
UR - https://europepmc.org/article/PMC/PMC13716863
AB - Background Predictive-processing accounts of BPD, including the hierarchical-prior-imbalance account proposed by Herzog, offer a framework for interpreting affective instability and impulsivity. The present task-fMRI study evaluates whether discriminative task-evoked patterns are consistent with this framework; because it does not directly estimate prediction errors, precision weighting, or hierarchical belief updating, it is not a mechanistic test of predictive processing.Methods To evaluate this question, this study developed a leakage-controlled multimodal machine learning framework for individualized classification of adolescents with BPD and healthy controls based on emotional Go/No-Go task-fMRI activity. Three feature domains were constructed: functional connectivity (FC), task-evoked general linear model (GLM)-derived temporal features, and clinical variables. Seven feature settings, including single-modality, dual-modality, and full multimodal combinations, were evaluated using multiple conventional machine learning classifiers, including logistic regression, support vector machine, k-nearest neighbors, Gaussian naive Bayes, multilayer perceptron, random forest, and XGBoost or gradient boosting.Results The results showed that task-evoked GLM-derived features provided the strongest single-modality discriminative information (AUC of 0.881 ± 0.055). SHAP analysis confirmed the predominance of GLM-derived features, accounting for 83.1% of total model attribution, compared to 9.5 and 7.5% for functional connectivity and clinical variables, respectively. At the neural level, discriminative features were distributed across visual/perceptual regions (calcarine, cuneus, occipital cortex), temporoparietal/default-mode hubs (angular and supramarginal gyri), frontal control regions (inferior frontal gyrus), and subcortical structures (putamen). This distributed pattern aligns with a hierarchical organization spanning perceptual input, contextual/self-referential processing, and top-down inhibitory control systems, consistent with the proposed model of hierarchical affective predictive imbalance.Conclusion Task-evoked GLM-derived features were the primary source of discriminative information for adolescent BPD, with functional connectivity and clinical variables providing modest additional value. The resulting classification pattern is consistent with hierarchical predictive-processing accounts but does not establish a specific predictive-processing mechanism. The same regional pattern may also reflect visual attention, salience processing, and executive inhibitory control.
SN - 1662-5161
ER -