Predicting Harm Scores from Patient Safety Event Reports.

Area of Science:

• Health Informatics
• Patient Safety Research
• Natural Language Processing

Background:

• The Agency for Healthcare Research and Quality's Harm Scale is standard in US hospitals for assessing patient safety event severity.
• Studies show moderate to poor inter-rater reliability for the Harm Scale, often attributed to subjective human judgment.
• Accurate harm severity identification is crucial for prioritizing safety analysis and interventions.

Purpose of the Study:

• To investigate the potential of automated text classification using narrative data from patient safety reports to objectively categorize harm scores.
• To reduce subjective human biases inherent in manual harm scoring.
• To enhance the efficiency and reliability of patient safety event analysis.

Main Methods:

• Utilized a corpus of patient safety reports from a US healthcare system.
• Evaluated various automated text classification algorithms to categorize harm scores based on narrative content.
• Focused on extracting and analyzing key information within reports to refine harm severity identification.

Main Results:

• Demonstrated the effectiveness and efficiency of automated text classification methods for categorizing harm scores.
• Showed a significant reduction in subjective human biases associated with harm score application.
• Confirmed that narrative data in patient safety reports contains crucial information for accurate harm severity assessment.

Conclusions:

• Automated text classification of patient safety reports offers a promising solution to improve harm score reliability and efficiency.
• The proposed methods can significantly mitigate human biases in patient safety event analysis.
• Findings support the development of semi-supervised tools to aid manual review and analysis of patient safety events.