Summary

Discovering associated injury patterns using machine learning

Abstract

Background

Overlooked injuries are common in the treatment of the acute trauma patient during primary and secondary assessment. Comprehensive identification of injury patterns may lead to earlier diagnosis and better treatment. Machine learning methods hold the promise to discover patterns on its own, referred to representation learning. We therefore asked: Can we use representation learning to identify correlated injury patterns in the acute trauma patient?

Methods

We included 3,208,681 patients from the Trauma Quality Improvement Program from 2016 through 2019. Of the included patients, the majority was male (61%), with a median age of 55 years (interquartile range 33-72). We carried out various representation learning methods (variational auto-encoders, VAE) to identify associated injury patterns per mechanism of injury and in patients with a mild, moderate and severe traumatic brain injury (as assessed with the Glasgow Coma Scale). The injury patterns found were evaluated using a novel developed body-spatial metric and the commonness of occurrence.

Results

Our implementation of a disentangled VAE reliably retrieved common non-obvious clusters, as evaluated with the body-spatial metric and clinical expert evaluation. In addition, the high-dimensional data was visualized in t-distributed stochastic neighbor embedding (t-SNE) plots.

Conclusions

The injury clusters found by applying representation learning methods can be used for educational purposes to inform trauma(-related) providers at tertiary assessment, academic and clinical purposes, and may improve future efforts leading to a recommender system for co-occurring injuries for decision support in the clinical workflow.