Towards Mining Electronic Health Records for Opioid ADE Surveillance. Yu, H, Brandt, C, Becker, W, & Kem, R In
abstract   bibtex   
Objectives: Prescription opioids are commonly used to treat acute and cancer-related pain, and, over the last two decades, have increasingly been used in the management of chronic non-cancer pain. Patients taking opioids can experience a wide range of adverse drug events (ADEs), including constipation, nausea/vomiting, pruritus, drowsiness and dizziness, hormonal dysfunction, depression, oversedation, falls, fractures, addiction, overdose, respiratory depression, sleep-disordered breathing, and death. Since such ADEs are frequently described in the unstructured electronic health record (EHR) notes, we are developing natural language processing (NLP) system to automatically extract opioid and ADEs from EHRs. The purpose of this study was to test out the feasibility of mining EHR notes for ADE detection using NLP approaches. Methods: We developed an annotation guideline using an interactive process during which physicians and linguists worked together to define rules and resolve discrepancy. Following the guideline, two annotators annotated 150 discharge summaries (or 8,672 sentences comprising 102,807 word tokens). The overall pairwise annotation agreement was 88%. The total number of annotated ADEs and medications were 103 and 3,290. Using this annotated corpus, we developed a NLP system to detect medication and ADE information. Our NLP system is trained on the supervised machine learning model Conditional Random Fields. We compared our NLP system with the state-of-the-art NLP system the MetaMap for ADE detection. Results: NLP performed well on discharge summaries on certain named entities, including frequency (92% F1), route (89% F1), dosage (87% F1), and medication (84% F1). Because the number of ADE instances is small, NLP performed poorly on ADE (24% F1). MetaMap performed on average 62% F1 for medication and 4% F1 for ADE. Implications: Our NLP system outperformed MetaMap for EHR notes ADE detection. NLP generally performs well with a sufficient size of annotated data. While the performance of ADE detection is low, more annotated data yielding a higher prevalence of ADEs would likely improve opioid ADE detection. Use of larger datasets is underway. Impacts: NLP has the potential to improve understanding of the nature and prevalence of opioid ADEs and, ultimately, advance the field of medication safety.
@inproceedings{ yu_towards_2015,
  title = {Towards Mining Electronic Health Records for Opioid {ADE} Surveillance},
  abstract = {Objectives:
Prescription opioids are commonly used to treat acute and cancer-related pain, and, over the last two decades, have increasingly been used in the management of chronic non-cancer pain. Patients taking opioids can experience a wide range of adverse drug events ({ADEs}), including constipation, nausea/vomiting, pruritus, drowsiness and dizziness, hormonal dysfunction, depression, oversedation, falls, fractures, addiction, overdose, respiratory depression, sleep-disordered breathing, and death. Since such {ADEs} are frequently described in the unstructured electronic health record ({EHR}) notes, we are developing natural language processing ({NLP}) system to automatically extract opioid and {ADEs} from {EHRs}. The purpose of this study was to test out the feasibility of mining {EHR} notes for {ADE} detection using {NLP} approaches.
 
Methods:
We developed an annotation guideline using an interactive process during which physicians and linguists worked together to define rules and resolve discrepancy. Following the guideline, two annotators annotated 150 discharge summaries (or 8,672 sentences comprising 102,807 word tokens). The overall pairwise annotation agreement was 88%. The total number of annotated {ADEs} and medications were 103 and 3,290. Using this annotated corpus, we developed a {NLP} system to detect medication and {ADE} information. Our {NLP} system is trained on the supervised machine learning model Conditional Random Fields. We compared our {NLP} system with the state-of-the-art {NLP} system the {MetaMap} for {ADE} detection.
 
Results:
{NLP} performed well on discharge summaries on certain named entities, including frequency (92% F1), route (89% F1), dosage (87% F1), and medication (84% F1). Because the number of {ADE} instances is small, {NLP} performed poorly on {ADE} (24% F1). {MetaMap} performed on average 62% F1 for medication and 4% F1 for {ADE}.
 
Implications:
Our {NLP} system outperformed {MetaMap} for {EHR} notes {ADE} detection. {NLP} generally performs well with a sufficient size of annotated data. While the performance of {ADE} detection is low, more annotated data yielding a higher prevalence of {ADEs} would likely improve opioid {ADE} detection. Use of larger datasets is underway.
 
Impacts:
{NLP} has the potential to improve understanding of the nature and prevalence of opioid {ADEs} and, ultimately, advance the field of medication safety.},
  eventtitle = {The 2015 {HSR}\&D/{QUERI} National Conference},
  author = {Yu, H and Brandt, C and Becker, W and Kem, R},
  date = {2015}
}

Downloads: 0