Classifying clinical notes with pain assessment using machine learning.

Area of Science:

• Medical Informatics
• Clinical Data Analysis
• Natural Language Processing

Background:

• Chronic pain affects millions, with documented deficits in pain care quality (PCQ).
• Existing methods lack intelligent approaches to identify PCQ indicators within electronic health records (EHR).
• Unstructured clinical notes in EHRs are an underutilized data source for pain assessment.

Purpose of the Study:

• To develop and validate a reliable classifier for identifying pain assessment indicators in unstructured EHR clinical notes.
• To analyze the documentation patterns of pain assessment subclasses in chronic pain patients.
• To leverage machine learning for enhanced pain assessment research using narrative EHR data.

Main Methods:

• Manual annotation of pain assessment qualifiers and descriptors in clinical notes.
• Development of a Random Forest classifier to detect pain assessment information.
• Utilized a dataset of 92 patients with chronic pain and 1058 clinical notes.

Main Results:

• The Random Forest classifier achieved high performance (accuracy 94%, PPV 95%, F1-score 94%, AUC 94%).
• Significant variations were found in documenting pain assessment subclasses, with pain site and intensity being most frequent.
• Etiology and aggravating factors of pain were documented in only 27% and 11% of notes, respectively.

Conclusions:

• Machine learning, specifically Random Forest, is effective for identifying pain assessment in clinical notes.
• This approach offers a novel method for analyzing pain assessment using unstructured EHR data, surpassing reliance on coded data alone.
• Improved identification of pain assessment indicators can contribute to better chronic pain management and enhanced pain care quality.