Bigger Is Not Always Better: Computational Efficiency in Lexical Prehospital Triage Modeling

Area of Science:

• Medical Artificial Intelligence
• Computational Efficiency in Healthcare
• Military Medicine

Background:

• Future large-scale combat operations necessitate efficient triage solutions for mass casualties.
• Artificial intelligence (AI) can enhance operational efficiency and enable triage at scale.
• Simpler machine learning models (MLMs) may offer advantages over large language models (LLMs) for edge implementation due to lower computational burden.

Purpose of the Study:

• To evaluate the clinical performance and computational efficiency of MLMs compared to LLMs for on-scene triage.
• To determine if computationally simpler AI can achieve equivalent clinical performance to complex AI in medical triage.

Main Methods:

• Retrospective study using data from a civilian critical care air transport service (2012-2021).
• Free-text clinician impressions and records of lifesaving interventions (LSIs) were abstracted.
• Two modeling approaches were used: MLMs (ensemble classifier with word frequencies) and LLMs (DistilBERT, BioBERT with tokenized word embeddings) to predict LSI receipt.

Main Results:

• MLMs and LLMs showed broadly equivalent performance in predicting LSIs (AUROC ~0.79-0.80).
• LLMs required substantially greater computational resources for both training and prediction compared to MLMs.
• Predictive equipoise was observed across seven independent LSI categories.

Conclusions:

• MLMs achieved equivalent clinical performance to LLMs for on-scene triage with a fraction of the computational resources.
• Nimble MLMs are advantageous for military AI, enabling deployment on edge devices and providing explainable decision support.
• Computational efficiency is critical for adopting and fielding AI medical capabilities in forward military environments.