How machine learning on real world clinical data improves adverse event recording for endoscopy

Area of Science:

• Gastroenterology
• Medical Informatics
• Machine Learning

Background:

• Endoscopic interventions are vital for diagnosing and treating gastrointestinal diseases.
• Accurate documentation is critical for patient safety and clinical outcomes.
• Adverse events following endoscopic procedures are often underreported.

Purpose of the Study:

• To evaluate a machine learning (ML) approach for detecting endoscopic adverse events.
• To systematically identify adverse events using real-world clinical metadata.
• To improve the accuracy of adverse event detection in endoscopy.

Main Methods:

• A random forest classifier was employed to detect adverse events: perforation, bleeding, and readmission.
• The model analyzed 2490 inpatient cases using structured hospital data, including ICD-codes and procedure timings.
• The ML approach utilized multiple metadata features for robust predictions.

Main Results:

• The ML model demonstrated significant improvements over baseline prediction accuracy for adverse events.
• Achieved AUC-ROC/AUC-PR values: 0.9/0.69 for perforation, 0.84/0.64 for bleeding, and 0.96/0.9 for readmissions.
• Highlighted the importance of multiple metadata features for effective prediction.

Conclusions:

• The semi-automated, privacy-preserving ML method effectively identifies documentation discrepancies and enhances quality control in endoscopy.
• This approach supports better clinical decision-making, quality improvement, and resource allocation.
• Reduces the risk of missed adverse events, thereby improving patient safety in endoscopic procedures.