Prediction of recidivism in a long-term follow-up of forensic psychiatric patients: Incremental effects of neuroimaging data

Abstract

One of the primary objectives in forensic psychiatry, distinguishing it from other psychiatric disciplines, is risk management. Assessments of the risk of criminal recidivism are performed on a routine basis, as a baseline for risk management for populations involved in the criminal justice system. However, the risk assessment tools available to clinical practice are limited in their ability to predict recidivism. Recently, the prospect of incorporating neuroimaging data to improve the prediction of criminal behavior has received increased attention. In this study we investigated the feasibility of including neuroimaging data in the prediction of recidivism by studying whether the inclusion of resting-state regional cerebral blood flow measurements leads to an incremental increase in predictive performance over traditional risk factors. A subsample ( N = 44) from a cohort of forensic psychiatric patients who underwent single-photon emission computed tomography neuroimaging and clinical psychiatric assessment during their court-ordered forensic psychiatric investigation were included in a long-term (ten year average time at risk) follow-up. A Baseline model with eight empirically established risk factors, and an Extended model which also included resting-state regional cerebral blood flow measurements from eight brain regions were estimated using random forest classification and compared using several predictive performance metrics. Including neuroimaging data in the Extended model increased the area under the receiver operating characteristic curve (AUC) from .69 to .81, increased accuracy from .64 to .82 and increased the scaled Brier score from .08 to .25, supporting the feasibility of including neuroimaging data in the prediction of recidivism in forensic psychiatric patients. Although our results hint at potential benefits in the domain of risk assessment, several limitations and ethical challenges are discussed. Further studies with larger, carefully characterized clinical samples utilizing higher-resolution neuroimaging techniques are warranted.

Forensic Brain-Reading and Mental Privacy in European Human Rights Law: Foundations and Challenges

“…An interesting example of such an 'interactive approach', is the study of Delfin et al, who performed a long-term follow-up study on the prediction of recidivism of forensic psychiatric patients, using both traditional risk assessment tools and resting-state SPECT data. One of the "incremental effects of neuroimaging data" they reported, was the increase of accuracy rate from 64% to 82% [21].…”

“…Brain data will likely be used in combination with other data, for example, to inform psychiatric evaluations. Given that epistemic access to other minds is in principle limited, and since current risk assessment tools are far from perfect [22], neuroimaging may make a significant contribution [19,21,23], especially with respect to people who refuse to cooperate, malinger, or simulateor are suspected of doing so.…”

Neuroprediction and A.I. in Forensic Psychiatry and Criminal Justice: A Neurolaw Perspective

“…Even more recently, a study by Delfin et al (2019) shows that improvements in recidivism prediction in forensic psychiatry might be possible by incorporating neuroimaging data into A.I. risk assessment models.…”

“…prediction model, containing neural measurements from eight brain regions, leads to an increase in predictive performance over traditional, empirical risk factors in a long-term follow-up of forensic psychiatric patients. Interestingly, they used 'classical' risk assessment combined with neuroimaging, which showed a better prediction in a forensic psychiatric population than the classical factors alone (Delfin et al, 2019).…”

“…neuroprediction may offer possibilities to decrease bias in risk assessment. However, also since neuroprediction may be incorporated in existing risk assessment tools (see the study by Delfin et al, 2019), bias will remain a problem as long as there is no solution to bias in algorithms in general.…”