Natural Language Processing-Based Visualization Framework for Adverse Events Extracted from Clinical Narratives: Towards Enhancing Clinical Interpretability

Abstract

Subjective adverse events (AEs), such as pain, are often under-recognized when relying solely on structured data. Although natural language processing (NLP) enables the extraction of AEs from narrative electronic health records (EHRs), interpretation of their temporal dynamics is difficult. Visualization methods can bridge this gap by transforming text-derived symptom data into clinically interpretable data. This study aimed to demonstrate the clinical value of a framework integrating NLP-based AE extraction with time-series visualization for otherwise invisible symptoms. Narrative texts, including progress notes, nursing records, and discharge summaries, were processed using MedNERN-CR-JA, a pretrained Japanese BERT-based model for entity recognition. AEs were visualized using Kaplan–Meier curves to represent the time to first onset and heatmaps, displaying all subsequent symptom documentation alongside supportive medication use. Among the 35042 eligible patients, 3094 received paclitaxel (PTX) and were matched to 3094 controls. PTX was associated with a higher risk of musculoskeletal symptoms (hazard ratio, 1.77; 95% confidence interval: 1.57–1.99). Kaplan–Meier curves showed earlier onset in PTX recipients, while heatmaps depicted recurrent documentation and the corresponding analgesic use. Restricting the analyses to the triweekly PTX regimen reduced the heterogeneity between inpatient and outpatient documentation and revealed a clearer alignment between the treatment cycles and acute symptom onset. This framework demonstrates the clinical value of visualizing NLP-extracted AEs from narrative EHRs, improving the resolution of subjective AE data, enhancing monitoring, and supporting patient-centered care and clinical decision making through complementary time-to-event and heatmap visualizations.