Real-time prediction of trauma-induced coagulopathy using an inverted transformer (trauma-former): a methodological feasibility and simulation study based on the ADEMP framework.

Abstract

Background: Trauma-induced coagulopathy (TIC) is a leading cause of preventable mortality following severe injury, yet prehospital recognition is fundamentally constrained by the limited sensitivity of static vital-sign metrics. Although fifth-generation (5G) wireless technology theoretically enables continuous biosignal transmission, robust computational frameworks for streaming risk prediction remain underdeveloped and largely unvalidated. To address this methodological gap, this study presents Trauma-Former, an inverted Transformer (iTransformer) architecture for real-time TIC prediction, trained and validated on high-fidelity synthetic physiological data within a rigorous in-silico engineering assessment framework.

Methods: An Ornstein-Uhlenbeck (OU) process generated 1 Hz vital-sign waveforms (heart rate [HR], systolic and diastolic blood pressure [SBP/DBP], SpO₂) for 1,240 simulated 30-minute transport episodes (50% TIC prevalence), structured according to the ADEMP framework. Trauma-Former embeds 60-second histories as independent variable tokens, applying self-attention across physiological streams to model inter-variable coupling. Models were benchmarked against an expanded set including a linear trend-based logistic regression (LR-trend), a 1D convolutional neural network (1D-CNN), a bidirectional gated recurrent unit (GRU), LSTM, XGBoost, PatchTST, Informer, and the shock index. An independent test set with 25% TIC prevalence assessed performance under clinically representative event rates. A sensitivity analysis incorporating binary missingness indicators probed the missing-completely-at-random (MCAR) assumption. Monte Carlo standard errors (MCSE) quantified simulation uncertainty throughout.

Results: On the balanced development set, Trauma-Former achieved an AUROC of 0.939 (95% CI 0.92-0.95; MCSE = 0.003) and an AUPRC of 0.88. Critically, the LR-trend model achieved an AUROC of 0.917 (95% CI 0.89-0.94), confirming that the synthetic task is predominantly solvable by detecting monotonic vital-sign drifts; the incremental contribution of iTransformer's cross-variable attention is consequently modest (AUROC gap: 0.022). Under a realistic 25% prevalence setting, AUROC remained 0.931 but positive predictive value (PPV) collapsed from 0.89 to 0.48-a 46% relative reduction-indicating that fewer than half of all model-generated alerts would correspond to true TIC events in a representative prehospital cohort. Adding binary missingness indicators yielded only a marginal PPV improvement of 0.02 under 30% MCAR data loss. The model demonstrated resilience to 30% packet loss and 4G-equivalent latency jitter with less than 5% AUROC degradation.

Conclusions: Trauma-Former provides a reproducible, open-source synthetic testbed for prehospital AI development. All reported metrics represent upper-bound estimates derived under deliberately simplified linear conditions. The severe PPV collapse under realistic prevalence identifies alarm fatigue as the paramount translational barrier and underscores that rigorous external validation on real-world prehospital data is the absolute prerequisite for any clinical application.