Assessing Revisit Risk in Emergency Department Patients: Machine Learning Approach

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems
• Machine Learning for Predictive Analytics

Background:

• Overcrowded emergency departments (EDs) can compromise care quality and staff workload.
• Assessing unscheduled return visits (URVs) is crucial for patient safety and cost reduction.
• Machine learning (ML) offers advanced tools for analyzing complex healthcare data.

Purpose of the Study:

• To develop an ML-assisted framework for identifying patients at high risk of ED revisits within 72 hours.
• To evaluate various ML models, feature sets, and encoding methods for optimal prediction.
• To enhance patient safety and reduce healthcare costs through proactive intervention.

Main Methods:

• Developed an ML system integrating structured electronic health records and unstructured clinical notes.
• Utilized a 5-year dataset of 184,687 ED visits from a tertiary medical center.
• Employed convolutional neural networks for feature extraction from narrative notes, combined with structured data.

Main Results:

• The ML framework achieved an area under the receiver operating characteristic curve of 0.705 and a recall of 0.718.
• Integrating unstructured physician notes with structured vital signs and demographics significantly improved predictive performance.
• The study demonstrated the effectiveness of combining diverse data sources for URV prediction.

Conclusions:

• An ML-assisted framework can serve as a valuable decision support tool for ED clinicians.
• Further exploration is warranted to refine the model for clinical implementation and patient care.
• The proposed framework shows potential for identifying high-risk patients for timely interventions.