Machine learning techniques for mortality prediction in emergency departments: a systematic review

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Artificial Intelligence in Healthcare

Background:

• Machine learning (ML) algorithms are increasingly explored for predicting patient outcomes.
• In-hospital mortality prediction in emergency departments (EDs) using vital signs is critical for timely intervention.
• Existing systematic reviews may not fully capture the performance and feasibility of these ML models.

Purpose of the Study:

• To systematically review the performance and clinical feasibility of ML algorithms for predicting in-hospital mortality in medical patients.
• To identify the types of ML models used and their reported performance metrics.
• To assess the quality and identify research gaps in studies predicting in-hospital mortality using vital signs in the ED.

Main Methods:

• A systematic literature search was conducted across major databases (Medline, Scopus, Embase) from 2010 to 2021.
• Included studies focused on ML models utilizing vital signs to predict in-hospital mortality for ED-admitted patients.
• Risk of bias was assessed using the prediction risk of bias assessment tool; critical appraisal followed standard checklists.

Main Results:

• Fifteen articles were included, evaluating eight different ML models (e.g., logistic regression, random forest, deep neural networks).
• Most studies exhibited a high risk of bias, particularly in statistical analysis, and often omitted details on data preprocessing and handling missing values.
• Nine articles lacked a specified time horizon for mortality prediction, impacting model interpretability.

Conclusions:

• The current body of research on ML for in-hospital mortality prediction in EDs has significant limitations, including high risk of bias and methodological inconsistencies.
• Eight recommendations are proposed to enhance the robustness and practical feasibility of ML models in clinical settings.
• Future research should focus on addressing identified gaps to develop reliable ML tools for early detection of patient deterioration.