FORMA Framework: Generating Clinical Vignettes while Preserving Cognitive Formulations

01 ABSTRACT

The study developed FORMA, which uses a cognitive model of PTSD to generate clinical vignettes conforming to specified cognitive structures. 16,500 vignettes were generated across 500 personas, 11 generation models, and three ablation conditions, and evaluated via external edge-recovery probes, clinical experts, a scaled LLM judge, and a user study with 100 licensed clinicians. Full-condition vignettes allowed recovery of the cognitive graph (MCC=0.41, AUC=0.70), whereas zero-shot generation was near chance (MCC=0.01, AUC=0.50). Experts rated full vignettes substantially higher, and clinicians perceived them as human-written 85% of the time vs. 22% for zero-shot. The authors argue that cognitive formulation can serve as an auditable specification for scalable synthetic clinical text.

02 KEY FINDINGS

1. Proposed FORMA framework for generating clinical vignettes based on cognitive models to ensure structural fidelity.
2. In PTSD vignettes, full-condition generation achieved significantly higher cognitive graph recoverability than zero-shot (MCC 0.41 vs. 0.01).
3. Expert ratings and clinician perceived authenticity were significantly higher than zero-shot.
4. 85% of clinicians mistakenly judged full-condition vignettes as human-written, compared to 22% for zero-shot.
5. FORMA reduced demographic disparity in perceived quality by 1.5 to 7 times.