Discharge-time prediction of 1-month posttraumatic stress symptom severity (PCL-5) after mechanical ventilation using a dual-attention 1D-CNN: Development and validation

Mechanical ventilation (MV) in the emergency intensive care unit (EICU) may expose patients to traumatic experiences that increase the risk of early post-discharge posttraumatic stress symptoms. A discharge-time tool for estimating symptom severity could support early risk stratification and follow-up planning. This single-center prospective observational cohort study enrolled mechanically ventilated EICU patients between July 17, 2025 and January 10, 2026, with 1-month follow-up. Pre-discharge demographic, clinical, and psychosocial variables were collected, and the Posttraumatic Stress Disorder Checklist for DSM-5 (PCL-5) assessments were completed at 1 month. After data cleaning, 500 participants were retained and randomly divided into a development set (n = 400) and an independent test set (n = 100). Based on univariate screening, clinical relevance, and data completeness, 11 predictors were entered into a dual-attention 1D-CNN multi-output regression model to predict the PCL-5 total score and four symptom-cluster scores, incorporating a constraint term enforcing the relationship between the total score and the sum of the cluster scores. Mean PCL-5 total was 38.0 ± 9.1; 70.8% of participants met the screening threshold for clinically significant posttraumatic stress symptoms (PCL-5 ≥ 33). The CNN-Attention model achieved R2 = 0.804 for total score on the test set (RMSE = 2.164; MAE = 1.935; MAPE = 7.43%) and R2 = 0.817–0.929 for clusters, outperforming comparator models. A discharge-time CNN-Attention model based on routinely available variables showed good performance in predicting early post-discharge PTSD symptom burden and symptom profiles after MV. Further multicenter studies are needed to externally validate the model and assess its clinical utility.