Adverse drug event presentation and tracking (ADEPT)

Adverse drug event presentation and tracking (ADEPT) detects potential adverse drug events in clinical notes using high-sensitivity natural language processing (NLP) to support pharmacovigilance and human adjudication.

Key Features:

• High-Sensitivity Detection: Employs high-sensitivity NLP algorithms to identify potential ADE mentions in unstructured clinical notes.
• Open-Source NLP Pipelines: Leverages open-source NLP pipelines to systematically scan clinical documentation for mentions of medications alongside signs and symptoms.
• Drug-Event Pair Extraction: Extracts and annotates drug-event pairs within the context of clinical narratives to support downstream review and adjudication.
• Support for Human Adjudication and Labeling: Generates candidate events for human review to enable creation of gold standard, patient-level labels for NLP-based pharmacovigilance.
• Refinement of Signal Sensitivity and Specificity: Combines automated detection with human oversight to improve distinction of true ADEs from symptoms related to comorbidities or other confounders.
• Efficiency in Human Review: Streamlines the review workflow, with a reported median review time of 89 seconds per patient in one study.

Scientific Applications:

• Pharmacovigilance with RWD: Monitors post-market drug safety using real-world data (RWD) extracted from clinical notes.
• Generation of Gold Standard Labels: Produces patient-level labeled data for training and evaluating NLP models in ADE detection.
• Signal Refinement and Validation: Supports validation of automated ADE signals through expert human adjudication to reduce false positives from confounding conditions.
• Pediatric Drug Safety Investigation: Enables investigation of potential ADEs in specific populations, exemplified by analysis of sildenafil-associated seizures in pediatric pulmonary hypertension patients.

Methodology:

Applies open-source, high-sensitivity NLP pipelines to scan clinical notes for mentions of medications and co-occurring signs and symptoms, extracts and annotates drug-event pairs, and presents candidate events for human review and adjudication; in a study of 149,029 notes from 982 pediatric pulmonary hypertension patients, ADEPT flagged potential sildenafil-associated seizures in 17% of NLP-flagged cases and human adjudication confirmed 0.96% as true ADEs, with a median review time of 89 seconds per patient.