# Unsupervised clustering analysis of trauma/non-trauma centers using hospital features including surgical care

## TL;DR

This study uses hospital data to identify patterns in surgical care delivery that distinguish trauma and non-trauma centers in Washington state.

## Contribution

The novel use of unsupervised clustering reveals surgical care patterns that go beyond traditional trauma center designations.

## Key findings

- Surgical care volume and variation distinguish hospitals in clustering analyses.
- Orthopedic procedures and factors like patient age and payer mix also influence cluster formation.
- Procedure volume aligns more closely with trauma center designation than procedure proportions.

## Abstract

Injuries are a leading cause of death in the United States. Trauma systems aim to ensure all injured patients receive appropriate care. Hospitals that participate in a trauma system, trauma centers (TCs), are designated with different levels according to guidelines that dictate access to medical and research resources but not specific surgical care. This study aimed to identify patterns of injury care that distinguish different TCs and hospitals without trauma designation, non-trauma centers (non-TCs).

We extracted hospital-level features from the state inpatient hospital discharge data in Washington state, including all TCs and non-TCs, in 2016. We provided summary statistics and tested the differences of each feature across the TC/non-TC levels. We then conducted 3 sets of unsupervised clustering analyses using the Partition Around Medoids method to determine which hospitals had similar features. Set 1 and 2 included hospital surgical care (volume or distribution) features and other features (e.g., the average age of patients, payer mix, etc.). Set 3 explored surgical care without additional features.

The clusters only partially aligned with the TC designations. Set 1 found the volume and variation of surgical care distinguished the hospitals, while in Set 2 orthopedic procedures and other features such as age, social vulnerability indices, and payer types drove the clusters. Set 3 results showed that procedure volume rather than the relative proportions of procedures aligned more, though not completely, with TC designation.

Unsupervised machine learning identified surgical care delivery patterns that explained variation beyond level designation. This research provides insights into how systems leaders could optimize the level allocation for TCs/non-TCs in a mature trauma system by better understanding the distribution of care in the system.