Natural Language Processing-Based Visualization Framework for Adverse Events Extracted from Clinical Narratives:...

Area of Science:

• Clinical informatics
• Biomedical data science
• Pharmacovigilance

Background:

• Subjective adverse events (AEs) are often under-recognized in structured electronic health record (EHR) data.
• Natural language processing (NLP) can extract AEs from narrative EHRs, but temporal analysis remains challenging.
• Visualization techniques can enhance the interpretability of text-derived AE data.

Purpose of the Study:

• To demonstrate the clinical utility of a framework integrating NLP-based AE extraction with time-series visualization for subjective symptoms.
• To improve the detection and monitoring of otherwise invisible AEs.
• To support patient-centered care and clinical decision-making.

Main Methods:

• Narrative EHR texts (progress notes, nursing records, discharge summaries) were processed using MedNERN-CR-JA, a Japanese BERT-based model for entity recognition.
• Adverse events (AEs) were visualized using Kaplan-Meier curves (time to first onset) and heatmaps (symptom documentation and supportive medication use).
• Analysis compared patients receiving paclitaxel (PTX) with matched controls.

Main Results:

• Paclitaxel (PTX) was associated with a significantly higher risk of musculoskeletal symptoms (HR, 1.77; 95% CI: 1.57-1.99).
• Kaplan-Meier curves indicated earlier onset of symptoms in PTX recipients.
• Heatmaps revealed recurrent symptom documentation and concurrent analgesic use, with clearer alignment to treatment cycles when focusing on triweekly PTX regimens.

Conclusions:

• The integrated NLP and visualization framework effectively enhances the resolution of subjective AE data from narrative EHRs.
• This approach improves AE monitoring, aiding clinical decision-making and patient-centered care.
• Visualizing time-to-event and symptom patterns offers valuable insights into drug-induced adverse events.