Multimodal imaging measures predict rearrest

Vaughn R. Steele, Eric D. Claus, Eyal Aharoni, Gina M. Vincent, Vince D. Calhoun, Kent A. KiehlFrontiers in Human Neuroscience 2015 · DOI

Abstract

Rearrest has been predicted by hemodynamic activity in the anterior cingulate cortex (ACC) during error-processing (Aharoni et al., 2013). Here, we evaluate the predictive power after adding an additional imaging modality in a subsample of 45 incarcerated males from Aharoni et al. (2013). Event-related potentials (ERPs) and hemodynamic activity were collected during a Go/NoGo response inhibition task. Neural measures of error-processing were obtained from the ACC and two ERP components, the error-related negativity (ERN/Ne) and the error positivity (Pe). Measures from the Pe and ACC differentiated individuals who were and were not subsequently rearrested. Cox regression, logistic regression, and support vector machine (SVM) neuroprediction models were calculated. Each of these models proved successful in predicting rearrest and SVM provided the strongest results. Multimodal neuroprediction SVM models with out of sample cross-validating accurately predicted rearrest (83.33%). Offenders with increased Pe amplitude and decreased ACC activation, suggesting abnormal error-processing, were at greatest risk of rearrest.

Introduction

Recidivism is a significant societal challenge, with roughly 68% of individuals in the U.S. rearrested within three years of release. Traditional methods for predicting reoffending often rely on static factors like criminal history or subjective clinical assessments. While previous research (Aharoni et al., 2013) demonstrated that hemodynamic activity in the Anterior Cingulate Cortex (ACC)—a brain region vital for error monitoring—could predict rearrest, it relied on a single imaging modality (fMRI). This paper investigates whether combining fMRI with Event-Related Potentials (ERPs) from EEG provides a more robust, multimodal window into the impulsive behaviors that lead to crime.

Methods

The researchers studied 45 incarcerated adult males using a Go/NoGo task, a classic measure of response inhibition and error processing. Participants were required to respond quickly to a "Go" stimulus but withhold their response for a "NoGo" stimulus.

The study utilized two primary neural measures:

  1. fMRI: To measure hemodynamic activity in the ACC during "False Alarms" (failed inhibition).
  2. ERP (EEG): To measure the ERN/Ne (Error-Related Negativity), an automatic signal occurring ~50ms after an error, and the Pe (Error Positivity), a later signal (200–400ms) associated with conscious error evaluation.

Three statistical approaches were compared to predict rearrest over a four-year follow-up: Cox proportional hazards regression, logistic regression, and Support Vector Machines (SVM).

Results and Neural Markers

The multimodal approach significantly improved prediction accuracy. The Support Vector Machine models provided the strongest results, achieving an overall accuracy of 78.05% and an 83.33% success rate in correctly identifying those who would be rearrested.

ERP waveforms and fMRI activation maps comparing rearrested (red) and non-rearrested (blue) groups.

A key finding was the predictive strength of the Pe component. Using Principal Component Analysis (PCA), the researchers found that "PC4" (representing the Pe component) was a massive predictor. In some models, every unit increase in PC4 amplitude corresponded to an 8.13-fold increase in the probability of rearrest. Interestingly, the Pe component—representing conscious error processing—was a much stronger predictor than the automatic ERN signal.

Discussion

The individuals at the highest risk exhibited a specific neural signature: increased Pe amplitude combined with decreased ACC activation. This "abnormal" error processing suggests a brain that is highly reactive to mistakes but lacks the regulatory control to change future behavior. Because fMRI tells us where the deficit is (the ACC) and ERP tells us what is happening in real-time, combining them outperformed models using only behavioral data or a single imaging type.

Conclusion

Neural measures remained significant predictors even after accounting for age and psychopathy scores, proving that brain data provides information that traditional behavioral assessments miss. Because these neural markers are dynamic, they could be targets for future interventions. For example, mindfulness training has been shown to modulate Pe amplitudes, potentially offering a way to "rewire" the brain to reduce recidivism risk.

Summary updated

Abstract

Rearrest has been predicted by hemodynamic activity in the anterior cingulate cortex (ACC) during error-processing (Aharoni et al., 2013). Here, we evaluate the predictive power after adding an additional imaging modality in a subsample of 45 incarcerated males from Aharoni et al. (2013). Event-related potentials (ERPs) and hemodynamic activity were collected during a Go/NoGo response inhibition task. Neural measures of error-processing were obtained from the ACC and two ERP components, the error-related negativity (ERN/Ne) and the error positivity (Pe). Measures from the Pe and ACC differentiated individuals who were and were not subsequently rearrested. Cox regression, logistic regression, and support vector machine (SVM) neuroprediction models were calculated. Each of these models proved successful in predicting rearrest and SVM provided the strongest results. Multimodal neuroprediction SVM models with out of sample cross-validating accurately predicted rearrest (83.33%). Offenders with increased Pe amplitude and decreased ACC activation, suggesting abnormal error-processing, were at greatest risk of rearrest.
BibTeX
@inproceedings{steele2015multimodal,
  author = {Vaughn R. Steele and Eric D. Claus and Eyal Aharoni and Gina M. Vincent and Vince D. Calhoun and Kent A. Kiehl},
  title = {{Multimodal imaging measures predict rearrest}},
  booktitle = {Frontiers in Human Neuroscience},
  year = {2015},
  doi = {10.3389/fnhum.2015.00425},
}
Figure