Quantifying risk factors in medical reports with a context-aware linear model.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Risk Assessment

Background:

• Electronic Health Records (EHRs) contain valuable clinical information.
• Quantifying mortality risk from textual data in EHRs is challenging due to contextual nuances.
• Existing methods struggle to accurately assess risk based on medical concept mentions.

Purpose of the Study:

• To develop a method for quantifying mortality risk associated with medical concepts in EHR text.
• To improve the accuracy of risk assessment by considering the textual context of medical concepts.
• To enhance the handling of rare medical concepts in risk quantification.

Main Methods:

• Proposed a multitask learning approach: context-aware linear modeling using regularized linear regression.
• Incorporated distributional similarity of concepts to improve performance on unseen or rare risk factors.
• Utilized a corpus of 531 EHR reports with 99,376 manually rated risk factors for evaluation.

Main Results:

• Context-aware linear modeling significantly outperformed single-task models.
• Integrating concept similarity further boosted performance, matching human annotator agreement levels.
• The model demonstrated high accuracy in quantifying risk factors within EHRs.

Conclusions:

• Automatic quantification of risk factors in EHRs can achieve human-level assessment performance.
• Multitask structure and handling of rare concepts are critical for accurate risk quantification.
• This approach offers a promising tool for leveraging EHR data for improved clinical risk assessment.