Abstract
Fabricating symptoms of Post-Traumatic Stress Disorder (PTSD) can hinder accurate clinical assessments via structured diagnostic interviews. Symptom simulation or fabrication is a known problem in PTSD assessments, with diverse motivations including unmet mental health issues, varied socio-economic factors and the potential for external gain from positive diagnoses. Here we introduce an artificial intelligence (AI) framework referred to as Algorithm VeRITAS (Vetting Response Integrity from cross-Talk in Adversarial Surveys), for detecting symptom fabrication in the context of PTSD diagnosis. In contrast to current approaches to fabrication detection which indirectly assess atypical symptom presentations, and have limited reliability, VeRITAS infers statistical dependencies inherent in true response patterns, flagging responses which violate these subtle constraints. With a study sample of n = 651 patients, VeRITAS has an Area Under the Curve (AUC) of $\ge 0.95 \pm 0.02$, with sensitivity > 95%, specificity > 88%, and positive likelihood ratio between 9.9 - 19.77. Additionally, VeRITAS is difficult-to-impossible to beat with coaching or training; we demonstrate that having advanced training in mental health diagnosis is not helpful in defeating the algorithm. Our tool offers an objective, disease-specific, fast (average time $\le$ 4 min) detection of simulated or feigned PTSD, and on wider adoption, can potentially help resources and disability concessions reach those genuinely in need, while helping to maintain integrity of clinical data. Moreover, reliably identifying patients who might be fabricating symptoms due to unmet mental health needs or socio-economic compulsions can ultimately improve outcomes in disadvantaged communities.
Introduction
The diagnosis of mental health disorders like PTSD relies heavily on self-reported symptoms. This dependency makes clinical assessments vulnerable to symptom fabrication—often termed malingering when motivated by secondary gain, such as financial incentives or legal settlements. Detecting fabricated PTSD is notoriously difficult. Traditional tools focus on identifying "atypical" symptoms, but these methods are time-consuming, require expert interpretation, and are highly susceptible to coaching. A subject familiar with DSM-5 criteria can often bypass these checks. There is a critical need for a fast, objective, and coaching-resistant tool to ensure the integrity of clinical data and the fair allocation of resources.
Methods: The VeRITAS Framework
The researchers developed VeRITAS (Vetting Response Integrity from cross-Talk in Adversarial Surveys). Unlike traditional methods that look for "fake" symptoms, VeRITAS employs a generative AI model to learn the subtle statistical dependencies—or "cross-talk"—between survey responses.
The system's core is a "Q-net" (Quasinet) composed of conditional inference trees. It models the joint distribution of responses from genuine PTSD patients. VeRITAS evaluates new response sequences based on two primary algorithmic properties:
- Complexity ($\kappa$): Genuine human responses exhibit a natural level of randomness. Fabricators often "over-structure" their answers to appear consistent, leading to sequences that are mathematically too simple or compressible. VeRITAS enforces a lower bound on this complexity.
- Surprise ($\nu$): Faked responses often violate the hidden correlations between survey items that the Q-net has learned. VeRITAS flags responses that exceed an upper bound of statistical "surprise."
To maintain robustness against memorization, the algorithm dynamically selects 20 items from a 211-item bank using SHAP values, creating approximately $10^{25}$ possible test variations.
Datasets and Evaluation
The framework was validated using three distinct datasets totaling 651 participants:
- VA Dataset: 304 US Veterans with clinician-confirmed CAPS-5 diagnoses, used to train the Q-nets on genuine response patterns.
- Prolific (PL) Dataset: 310 healthy individuals instructed to fake PTSD symptoms as convincingly as possible.
- Mental Health Professionals (PS) Cohort: 27 doctorate-level professionals (psychiatrists, psychologists, forensic specialists) instructed to use their expertise to defeat the algorithm.
Key Results
VeRITAS demonstrated exceptional performance across all cohorts. It achieved a lower bound AUC of $\ge 0.95 \pm 0.02$. At a 20% population prevalence of fabrication, the system maintains 94.2% sensitivity and 93% specificity.
Notably, the algorithm proved highly resistant to domain expertise. Only 3.7% of mental health professionals successfully faked PTSD without being flagged—a rate even lower than the general population cohort (5.8%). This suggests that clinical knowledge of PTSD criteria does not help a person mimic the complex statistical dependencies learned by the AI. Furthermore, the assessment is extremely efficient, with an average completion time of just 3.4 minutes.
Discussion
While VeRITAS offers a powerful tool for maintaining clinical integrity, its deployment requires careful ethical consideration. False positives could lead to the denial of care for genuine patients. The authors emphasize that VeRITAS should serve as a flag for human review rather than an automated gatekeeper. Future work will focus on expanding the model to detect other reporting styles, such as defensive under-reporting or careless responding, and ensuring the model remains unbiased across culturally diverse populations.
Summary updated