Text preprocessing for improving hypoglycemia detection from clinical notes - A case study of patients with diabetes

Area of Science:

• Medical Informatics
• Clinical Data Analysis
• Diabetes Mellitus Management

Background:

• Hypoglycemia is a critical safety concern in diabetes management.
• Electronic medical records are valuable for hypoglycemia detection, but ICD codes have limitations.
• Existing methods struggle with accurate and sensitive identification of hypoglycemia events.

Purpose of the Study:

• To develop and evaluate text preprocessing methods for improved automatic detection of hypoglycemia from clinical notes.
• To enhance the accuracy and sensitivity of identifying hypoglycemia events in electronic health records.
• To address the limitations of ICD codes in capturing real-time hypoglycemia occurrences.

Main Methods:

• Developed three text preprocessing techniques: stop word filtering, medication signaling, and ICD narrative enrichment.
• Utilized clinical notes from the VA Maryland Healthcare System, selected based on diabetes, hypoglycemia codes, glucose levels, and text references.
• Validated methods on datasets from 2009 and 2014, with manual review by physician judges to confirm hypoglycemia presence.

Main Results:

• All proposed preprocessing methods significantly increased the F1 score for hypoglycemia detection (5.3–7.4% and 7.7–9.4% in different datasets).
• Stop word filtering yielded the largest individual performance improvement (up to 7.4%).
• Combining methods demonstrated synergistic gains, particularly in the 2014 dataset, though not always yielding further improvement over combined methods.

Conclusions:

• Text preprocessing methods substantially enhance hypoglycemia detection accuracy from clinical notes.
• Stop word filtering is a key contributor to performance improvement.
• ICD narrative enrichment specifically boosts the recall of hypoglycemia detection, leading to more comprehensive identification.