High-accuracy prediction of mental health scores from English BERT embeddings trained on LLM-generated synthetic...

To assess whether synthetic-only first-person clinical self-reports generated by a large language model (LLM) can support accurate prediction of standardized mental-health scores, enabling a privacy-preserving path for method development and rapid prototyping when real clinical text is unavailable.

We prompted an LLM (Gemini 2.5; July 2025 snapshot) to produce English-language first-person narratives that are paired with target scores for three instruments-PHQ-9 (including suicidal ideation), LSAS, and PCL-5. No real patients or clinical notes were used. Narratives and labels were created synthetically and manually screened for coherence and label alignment. Each narrative was embedded using bert-base-uncased (mean-pooled 768-d vectors). We trained linear/regularized linear (Linear, Ridge, Lasso) and ensemble models (Random Forest, Gradient Boosting) for regression, and Logistic Regression/Random Forest for suicidal-ideation classification. Evaluation used 5-fold cross-validation (PHQ-9/SI) and 80/20 held-out splits (LSAS/PCL-5). Metrics: MSE, $R^2$ , MAE; classification metrics are reported for SI.

Within the synthetic distribution, models fit the label-text signal strongly (e.g., PHQ-9 Ridge: MSE $4.41\pm 0.56$ , $R^{2}0.92\pm 0.02$ ; LSAS Gradient Boosting test: MSE $75.00$ , $R^{2}0.95$ ; PCL-5 Ridge test: MSE $35.62$ , $R^{2}0.85$ ).

LLM-generated self-reports encode a score-aligned signal that standard ML models can learn, indicating utility for privacy-preserving, synthetic-only prototyping. This is not a clinical tool: results do not imply generalization to real patient text. We clarify terminology (synthetic text vs. real text) and provide a roadmap for external validation, bias/fidelity assessment, and scope-limited deployment considerations before any clinical use.