A deep learning-based model combining quantitative parameters of initial chest computed tomography and clinical variables for the prediction of posttraumatic acute respiratory distress syndrome

Background: Acute respiratory distress syndrome (ARDS) frequently complicates severe trauma, and early risk stratification for this condition remains challenging. This study aimed to develop and internally validate a combined prediction model integrating deep learning-based quantitative parameters of initial chest computed tomography (CT) with clinical variables for the early prediction of posttraumatic ARDS, with a focus on automation and interpretability for potential integration with emergency workflows.

Methods: A total of 173 patients with trauma admitted to West China Hospital of Sichuan University from January to June 2024 were retrospectively screened, and 105 patients were finally enrolled (34 in the ARDS group and 71 in the nonARDS group). Clinical data were collected, and quantitative parameters (e.g., opacity score, opacity volume, opacity percentage, highattenuation opacity volume, and mean CT value) were automatically extracted with deep learning-based CT quantitative analysis software. Univariate and multivariate logistic regression analyses were performed to identify factors associated with ARDS. The final multivariate model was expressed as follows: logit(ARDS risk) = −4.623 + [0.142 × Injury Severity Score (ISS)] + [0.247 × procalcitonin (PCT)] + (0.113 × bilateral lung opacity percentage). Model stability was evaluated via bootstrap resampling (1,000 repetitions), and the predictive performance was assessed according to the receiver operating characteristic (ROC) curves.

Results: Compared with the non-ARDS group, the ARDS group had a higher ISS (40 vs. 25, P<0.001), lactate level (3.9 vs. 1.9 mmol/L, P<0.001), and PCT level (3.2 vs. 0.2 ng/mL, P<0.001), as well as a lower Glasgow Coma Scale (GCS) score (12 vs. 15, P<0.001) and SpO₂ (94.8% vs. 96.7%, P=0.038). The ARDS group also had a significantly higher bilateral lung opacity volume (186.3 vs. 56.7 mL, P=0.001), opacity percentage (5.2% vs. 1.8%, P=0.002), and high-attenuation opacity volume (45.8 vs. 9.0 mL, P=0.001). Multivariate analysis identified the independent predictors of ARDS to be ISS [odds ratio (OR) =1.15; 95% confidence interval (CI): 1.06–1.24; P<0.001], PCT (OR =1.28; 95% CI: 1.04–1.58; P=0.02), and bilateral lung opacity percentage (OR =1.12; 95% CI: 1.04–1.21; P=0.004). The combined model achieved an area under the curve (AUC) of 0.892 (95% CI: 0.832–0.952). At the optimal cutoff of 0.385, the sensitivity and specificity were 82.4% and 85.9%, respectively. Bootstrap internal validation confirmed good model stability (mean AUC =0.878; 95% CI: 0.821–0.934).

Conclusions: Quantitative parameters of initial chest CT within 24 hours after trauma, especially bilateral lung opacity percentage, are associated with the development of ARDS in patients with trauma and may serve as objective imaging biomarkers for early risk assessment. The prediction model combining ISS score and PCT with automated CT quantitative parameters demonstrated good discriminative ability (AUC =0.892) and clinical operability and can be automated within minutes in a research setting. Further implementation studies are needed before the model can be implemented after the initial CT scan in clinical settings; nonetheless, the model may serve as a decision support tool and as a component in implementing early lung-protective strategies.