Artificial intelligence-driven clustering for phenotyping life-threatening prehospital trauma

TL;DR

This study uses AI to identify three trauma patient groups with very different mortality risks, which could improve emergency care and resource allocation.

Contribution

A novel AI-driven clustering method was used to phenotype prehospital trauma patients based on mortality risk and injury patterns.

Findings

Three distinct trauma phenotypes were identified with mortality rates of 93.1%, 68.1%, and 10.6%.

Cluster T-1 was primarily associated with traumatic brain injuries, thoracic trauma, and burns.

Cluster T-3 predominantly involved orthopedic trauma with the lowest mortality rate.

Abstract

Traumatic patients usually suffer from several complex conditions that hinder their risk characterization. The aim of this study was to derive phenotypes of prehospital acute life-threatening trauma via nonsupervised artificial intelligence (AI) clustering methods.

This was a prospective multicenter study in adult trauma patients treated in prehospital care and transferred to the emergency department. The study included 147 ambulances, 4 helicopters, and 11 hospitals in Spain between 1 January 2021 and 31 August 2024. Epidemiological variables, trauma-related data, baseline vital signs and blood tests were collected. The primary outcome was all-cause 2-day in-hospital mortality.

A total of 1474 patients were included, with a 2-day in-hospital mortality rate of 8.3%. The selected clustering method identified three clusters: the T-1 phenotype comprised 6.9% (101 cases) with a mortality rate of 93.1%, the T-2 phenotype represented 23.6% (348 cases) with a mortality rate of 68.1%, and T-3 represented 69.5% (1,025 cases) with a mortality rate of 10.6%. The T-1 phenotype mainly involves traumatic brain injuries, followed by thoracic trauma and burns; the T-2 phenotype presents a similar distribution; and the T-3 phenotype predominantly involves orthopedic trauma.

The AI method identified three clusters with implications for therapy and outcomes. This novel approach could help emergency medical services characterize trauma patients by providing benefits, treatment and resource optimization.

The online version contains supplementary material available at 10.1186/s13049-026-01553-0.

Do prehospital phenotypes characterize trauma patients’ risk of worsening and mortality?

Unsupervised artificial intelligence methods revealed three phenotypes. Cluster #1 presented a mortality rate of 93.1%, which was associated mainly with traumatic brain injuries, followed by thoracic trauma and burns. Cluster #2 had a 68.1% mortality rate with a similar distribution of traumatic causes. Cluster #3 presented a mortality rate of 10.6%, involving orthopedic traumas.

Our results provide specific phenotypes, which could help to improve patients’ care and optimize resources.

The online version contains supplementary material available at 10.1186/s13049-026-01553-0.

Figures