Trust Your Neighbors: Multimodal Patient Retrieval for TBI Prognosis.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Neurotrauma Research

Background:

• Accurate triage of traumatic brain injury (TBI) is crucial for patient outcomes.
• Emergency department (ED) disposition decisions for head injuries are often inconsistent.
• Existing decision support frameworks for TBI are limited.

Purpose of the Study:

• To introduce RAPID-TBI, a multimodal system for predicting ED disposition in TBI patients.
• To enhance clinical decision-making through example-based retrieval and emulate case-based reasoning.
• To improve the consistency and trustworthiness of TBI care.

Main Methods:

• Developed RAPID-TBI, a multimodal system integrating head CT scans, reports, exam findings, labs, vitals, and demographics.
• Utilized an attention-based encoder to generate patient embeddings for disposition classification.
• Employed example-based retrieval to mimic clinical case-based reasoning for enhanced interpretability.

Main Results:

• RAPID-TBI achieved state-of-the-art classification performance in predicting ED disposition.
• The system demonstrated consistent performance and resilience across institutional and temporal generalizability.
• Explored small language models for prompt-based, retrieval-guided prediction without fine-tuning.

Conclusions:

• RAPID-TBI offers a promising approach to trustworthy, personalized decision support in TBI care.
• The system provides consistent, individualized, and clinically grounded predictions.
• This multimodal, example-based retrieval system enhances interpretability in TBI disposition decisions.