Interpretable Multiomics Models for Predicting Surgical Interventions and Blood Transfusion Requirements in Traumatic...

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Predicting surgical and transfusion needs in traumatic brain injury (TBI) patients is difficult in emergency situations.
• Current methods lack accuracy and timeliness for effective intervention.

Purpose of the Study:

• To develop and validate multiomics data fusion (MDF) models for predicting surgical intervention and blood transfusion requirements in TBI patients.
• To assess the generalizability and practical utility of these models in multicenter emergency settings.

Main Methods:

• Integrated clinical biomarkers, neural radiological imaging, and clinical text mining using MDF.
• Developed and validated predictive models across four multicenter cohorts (N=2219).
• Employed SHapley additive exPlanations (SHAP) to identify key predictive features.

Main Results:

• MDF models predicted surgical needs a median of 3 hours before intervention, outperforming single-domain approaches (F1 scores: 0.63-0.85).
• Transfusion models showed strong cross-center performance (F1 scores: 0.78, 0.74) and correlated with transfusion volumes (R=0.687, R=0.580).
• Radiological features drove surgical predictions; clinical parameters (lactate, GCS, pupillary reflex, hemoglobin) drove transfusion predictions. A simplified model maintained high performance (AUC: 0.81, 0.75).

Conclusions:

• MDF models demonstrate cross-center generalizability and practical utility for predicting TBI patient needs in emergency settings.
• These models can support clinical decision-making and improve TBI patient management.
• The integration of multiomics data offers a powerful approach for complex clinical predictions.