Inter hospital external validation of interpretable machine learning based triage score for the emergency department...

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Health Informatics

Background:

• Emergency departments (ED) face challenges in prioritizing patients with limited resources.
• Machine learning (ML) tools can aid in emergency triage.
• Previous ML-based triage tools were developed using single-center data.

Purpose of the Study:

• To develop and validate a multicenter, interpretable ML model for predicting 2-day mortality in emergency department patients.
• To assess the performance of the Score for Emergency Risk Prediction (SERP) across different Korean hospital cohorts using a common data model (CDM).
• To enable multicenter research without direct data sharing.

Main Methods:

• Retrospective cohort study including adult emergency visits from 3 Korean hospitals (2016-2017).
• Utilized a common data model (CDM) for standardized multicenter data analysis.
• Developed and validated an interpretable ML model (SERP) for 2-day mortality prediction with inter-hospital validation.

Main Results:

• The study included over 225,000 emergency visits across three hospitals.
• The 2-day mortality rates ranged from 0.51% to 0.65%.
• Inter-hospital validation demonstrated high accuracy, with AUROC values of at least 0.899 (0.858-0.940).

Conclusions:

• A multicenter, interpretable ML model (SERP) was successfully developed for predicting 2-day mortality in ED patients using a CDM.
• The model achieved high accuracy in inter-hospital external validation, indicating its generalizability.
• This approach facilitates robust multicenter research and tool development without compromising data privacy.