Machine-learning classification of documented pedophilic sexual interest from eye-tracking patterns.

Assessment of pedophilic sexual interest is often embedded in crucial clinical and forensic contexts and could therefore benefit from non-invasive measures that complement clinical interviews and self-report. We investigated whether multivariate eye-tracking and behavioral patterns could classify documented pedophilic sexual interest at the individual level. The discovery sample comprised 124 adult men from five German sites, including 34 with and 90 without a documented pedophilia diagnosis according to ICD-10. Participants completed the Initial Orientation Task (IOT), Sexual Distractor Task (SDT), and Sexual Stimulus Rating task. Linear support vector machines were trained within a repeated nested cross-validation framework to predict pedophilia strictly on these eye-tracking features. The best-performing main model combined IOT and SDT, achieving 88.01% balanced accuracy, 88.24% sensitivity, 87.78% specificity, and an area under the curve of 0.93 (95% confidence interval, 0.87-0.99). Exploratory sparse modeling using only five IOT and four SDT features still yielded 87.65% balanced accuracy. This sparse model further classified all 32 participants in an independent healthy-control cohort as negative (100% specificity). Predictive features reflected relative attention to child versus adult stimuli and allocation of attention between distractors and cognitive targets. These findings support further investigation of eye-tracking measures as potential adjuncts to clinical assessment, in potentially high-stakes clinical and forensic decision-making. Prospective validation in representative cohorts containing both positive and negative cases is required to further establish generalizability and clinical utility.