Introduction: Individualized treatment prediction is crucial for the development and selection of personalized psychiatric interventions. Here, we use random forest classification via pretreatment clinical and demographical (CD), functional, and structural magnetic resonance imaging (MRI) data from patients with borderline personality disorder (BPD) to predict individual treatment response.
Methods: Before dialectical behavior therapy (DBT), 31 female patients underwent functional (three different emotion regulation tasks) and structural MRI. DBT response was predicted using CD and MRI data in previously identified anatomical regions, which have been reported to be multimodally affected in BPD.
Results: Amygdala and parahippocampus activation during a cognitive reappraisal task (in contrasts displaying neural activation for emotional challenge and for regulation), along with severity measures of BPD psychopathology and gray matter volume of the amygdala, provided best predictive power with neuronal hyperractivities in nonresponders. All models, except one model using CD data solely, achieved significantly better accuracy (>70.25%) than a simple all-respond model, with sensitivity and specificity of >0.7 and >0.7, as well as positive and negative likelihood ratios of >2.74 and <0.36 each. Surprisingly, a model combining all data modalities only reached rank five of seven. Among the functional tasks, only the activation elicited by a cognitive reappraisal paradigm yielded sufficient predictive power to enter the final models.
Conclusion: This proof of principle study shows that it is possible to achieve good predictions of psychotherapy outcome to find the most valid predictors among numerous variables via using a random forest classification approach.
Keywords: functional MRI; machine learning; multimodal data analysis; prediction of treatment response; random forest; structural MRI.
© 2019 The Authors. Brain and Behavior published by Wiley Periodicals, Inc.