Using Explainable Artificial Intelligence Models (ML) to Predict Suspected Diagnoses as Clinical Decision Support.

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Health Informatics

Background:

• Emergency departments (EDs) face increasing complexity and patient volumes.
• Shortages of specialist medical staff in EDs can lead to diagnostic errors and adverse outcomes.
• High time expenditure and costs are associated with current diagnostic processes in emergency care.

Purpose of the Study:

• To develop and evaluate explainable artificial intelligence (XAI) models for diagnosing common emergency conditions.
• To leverage the German national registry of medical emergency departments (AKTIN-registry) for robust model training.
• To create a portable clinical decision support system (CDSS) to enhance ED diagnostic accuracy and patient outcomes.

Main Methods:

• Analysis of 137,152 patient samples and 51 features (vital signs, symptoms) from the AKTIN-registry.
• Development of XAI models, including logistic regression (LR) and random forest (RF), for the 20 most frequent diagnoses.
• Evaluation of model performance using metrics such as area under the curve (AUC) and predictive accuracy.

Main Results:

• XAI models demonstrated high performance with AUC values of 0.98 (LR) and 0.99 (RF).
• Predictive accuracies reached 0.927 for LR and 0.99 for RF.
• The study identified top-performing models suitable for integration into a CDSS.

Conclusions:

• Explainable AI models show significant potential to improve diagnostic accuracy in emergency departments.
• The developed models can aid in reducing diagnostic errors, improving patient outcomes, and lowering healthcare costs.
• A portable CDSS integrating the best XAI model will be clinically validated in German EDs.