Machine learning in diagnostic support in medical emergency departments

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Clinical decision support systems

Background:

• Emergency department (ED) diagnosis is complex and increasingly challenged by aging populations.
• Diagnostic errors and patient logistics require innovative solutions.
• Machine learning (ML) offers potential for assisting physicians in high-pressure ED environments.

Purpose of the Study:

• To assess the feasibility of training ML algorithms to support ED physicians.
• To evaluate the diagnostic accuracy and predictive capabilities of these algorithms.
• To determine the impact of ML-assisted diagnostics on patient outcomes and resource utilization.

Main Methods:

• A cohort study of 9,190 ED admissions across two hospitals.
• Training 19 ML algorithms on 80% of patient data (biochemical, nurse-registered data) for 19 outcomes.
• Validation of algorithms on the remaining 20% of patient data.

Main Results:

• ML algorithms achieved high predictive accuracy for outcomes like 7-day mortality (AUC 91.4%) and 30-day mortality (AUC 91.3%).
• Algorithms demonstrated strong performance in predicting safe discharge (AUC 87.3%).
• Implementation of the biochemical analysis list reduced subsequent venipunctures by 22%.

Conclusions:

• ML algorithms can be developed with high Area Under the Curve (AUC) for ED applications.
• These algorithms show potential to reduce diagnostic errors and improve patient logistics.
• The study successfully demonstrated the feasibility of reducing venipunctures through ML-guided analysis.