B – 149 Data-Driven Detection of Symptom Exaggeration in Pain Patients Using Unsupervised Learning

Abstract Objective This study tested the hypothesis that distinct patterns of symptom reporting—valid, invalid, and random—could be empirically identified using unsupervised learning, specifically k-means cluster analysis. The analysis evaluated whether response profiles could be reliably classified when varying proportions of simulated random responders were introduced into a large normative dataset. Method The sample included 12,948 personal injury claimants (51% male, 49% female; mean age = 42, SD = 12) undergoing rehabilitation or medicolegal evaluations. Participants completed the Pain Symptom Report (PSR), a 60-item self-report measure using a 6-point scale. Eight validity scales were developed based on empirical response patterns: Infrequent Symptoms, Common Symptoms, Frequent Symptoms, Minimizing, Neutral Responding, Maximizing, Extreme Responding, and Inconsistent Responding. Iterative k-means cluster analyses (k = 2–20) were conducted on datasets augmented with 125, 250, 500, or 1,000 computer-generated random responders. Each clustering solution was evaluated for model convergence, cluster purity, and classification accuracy. Results Across all conditions, the random responder group was consistently and reliably identified with high accuracy (Average Balanced Accuracy = 96.2%). The optimal 9-cluster solution revealed empirically distinct response profiles: Random Responders, Symptom Magnifiers, Extreme Symptom Magnifiers, Hybrid Responders, Symptom Minimizers, Guarded/Inconsistent Minimizers, Neutral Responders, Broad Symptom Endorsers, and Average Responders. Each group showed unique score patterns, with classification accuracy ranging from 96.1% to over 99% across clusters. Conclusion PSR validity scales effectively distinguish between valid and invalid response styles. These findings support the value of unsupervised learning in empirically classifying symptom validity in both clinical and forensic assessment contexts.