Modeling PTSD Trajectories with Conditional SVAEs and Syntetic Data Generation: Data-Efficient Prediction and Outcome-Specific Explainability

Childhood trauma initiates complex psychiatric trajectories, but predictive modeling is hampered by data scarcity. We ask if a conditional Transformer-SVAE can learn patient embeddings from longitudinal surveys to improve PTSD prediction and reveal clinical drivers of model performance. Our framework uses a conditional SVAE to generate synthetic patient trajectories, addressing class imbalance. In our experiments, combining real and synthetic data \textbf{increased the identification of true positive PTSD cases by 82\%} over a real-data-only baseline, achieving a top F1-score of 0.683. Ablation studies confirm that architectural choices like "free bits" are essential for generating effective augmentation data. Finally, by stratifying SHAP explanations by outcome (TP, FP, FN, TN), we transform interpretability into a diagnostic tool, revealing how the model's reasoning differs between correct and incorrect predictions. This allows for targeted clinical insights, such as identifying when the model over-weights hopelessness signals, making predictions more transparent and clinically actionable.