Children referred as victims and alleged offenders for forensic psychiatric evaluation: distinct psychiatric profiles and machine-learning–based classification of post-traumatic stress disorder in a Turkish sample

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Children referred for forensic psychiatric evaluation represent a heterogeneous and vulnerable population comprising both victims and alleged offenders, who may differ substantially in their psychopathological and cognitive characteristics. Evidence from non-Western legal and cultural contexts remains limited, and few studies have applied complementary classical and machine-learning approaches to identify factors associated with post-traumatic stress disorder (PTSD) in this group. This study aimed to compare sociodemographic, clinical, and cognitive characteristics between victims and alleged offenders and to identify factors associated with PTSD. Methods This retrospective cross-sectional study reviewed forensic psychiatric evaluation records of 420 children and adolescents, including 104 alleged offenders and 316 victims. Factors associated with PTSD were examined using logistic regression and complementary supervised machine-learning models with SHAP-based interpretation. Results Alleged offenders were significantly older and predominantly male, with higher rates of special education support, school dropout, medical comorbidity, attention-deficit/hyperactivity disorder (ADHD), conduct disorder (CD), intellectual disability (ID), and specific learning disorder (SLD). Victims showed significantly higher rates of PTSD (19.4% vs. 1.0%) and major depressive disorder (MDD) (15.8% vs. 2.0%). In the multivariable model, victim status (adjusted OR = 10.59, 95% CI 1.98–196.58), female sex (adjusted OR = 2.47, 95% CI 1.07–6.12), and current psychiatric treatment (adjusted OR = 43.10, 95% CI 12.73–269.56) were independently associated with higher odds of PTSD, whereas special education support was associated with lower odds (adjusted OR = 0.27, 95% CI 0.06–0.87). All machine-learning models achieved area under the receiver operating characteristic curve (AUC) values exceeding 0.80, with Gradient Boosting yielding the highest discrimination (AUC = 0.860); SHAP analysis identified psychiatric diagnosis and acute stress disorder as the most influential classification features. Conclusions Victims and alleged offenders referred for forensic evaluation present with distinct internalizing and externalizing profiles. Victim status, female sex, and current psychiatric treatment were independently associated with PTSD, and machine-learning analyses were consistent with a substantial contribution of psychiatric morbidity. These findings support differentiated, profile-specific assessment of children within the Turkish forensic system and indicate the value of complementary analytic approaches.