ODD: A Benchmark Dataset for Opioid-Related Aberrant Behavior Detection in Natural Language Processing
Core Concepts
ORAB detection dataset ODD aims to identify aberrant behaviors and opioid-related information from EHR notes.
Abstract
Introduction of ODD, a biomedical NLP benchmark dataset for ORAB detection.
Explored state-of-the-art NLP models for ORAB identification.
Performance analysis shows the effectiveness of prompt-based fine-tuning.
Error analysis reveals confusion between confirmed and suggested aberrant behaviors.
Socio-demographic analysis indicates age impacts performance.
Prospective social impact discussed with ethical considerations.
Limitations include data source constraints and language limitations.
Future work involves improving performance in uncommon categories through advanced NLP approaches and data augmentation with LLMs.
Directory:
Introduction to ODD and ORAB detection task
Novel benchmark dataset introduced for ORAB detection in EHR notes.
State-of-the-art NLP Models Experimentation
Explored fine-tuning and prompt-based fine-tuning models for ORAB identification.
Performance Analysis
Prompt-based fine-tuning outperformed traditional fine-tuning, especially in uncommon categories.
Error Analysis
Confusion between confirmed and suggested aberrant behaviors identified as a common error.
Socio-demographic Analysis
Age group analysis showed differences in performance based on age categories.
Prospective Social Impact & Ethical Considerations
Positive impact on opioid abuse prevention discussed along with potential negative impacts on patient autonomy.
Limitations & Future Work
Data source limitations, language constraints highlighted, future work includes improving performance through advanced NLP methods.
ODD
Stats
"Experimental results show that the best model achieved the highest 88.17% on macro average area under precision recall curve."
"Among 331,794 EHR notes of 299,712 patients in MIMIC-IV database, approximately 57% of patients were prescribed opioids during their hospitalization."