PaperlayerView exact revision on arXiv

Source-linked AI summary

Natural Language Processing for EHR-Based Computational Phenotyping

Zexian Zeng, Yu Deng, Xiaoyu Li, Tristan Naumann, Yuan Luo

arXiv:1806.04820v2cs.CL

TL;DR

NLP-based computational phenotyping addresses the challenge of extracting clinically meaningful phenotypes from heterogeneous EHR data for research applications. This review surveys keyword and rule-based systems, supervised and unsupervised learning, deep learning, and multimodal approaches. It finds no dominant method: simpler systems can perform well but scale poorly, while newer methods offer performance or discovery advantages with continuing interpretability and generalizability challenges.

  • Problem

    Computational phenotyping must extract meaningful phenotypes from structured EHR data and unstructured clinical narratives, but existing approaches involve trade-offs in manual effort, labeling, scalability, and generalizability.

  • Method

    The paper reviews NLP-based computational phenotyping methods spanning keyword search, rule-based systems, supervised and unsupervised learning, deep learning, and combined data modalities.

  • Results

    The review finds no dominant method: keyword and rule-based systems often achieve high accuracy, supervised methods are widely used, and deep or unsupervised methods offer performance or phenotype-discovery advantages.

  • Takeaways & Limitations

    Method selection remains domain-dependent, while combining heterogeneous data sources and modalities is an important direction for improving computational phenotyping.

  • Takeaways & Limitations

    Rule adaptation remains labor-intensive and limits generalizability, while deep-learning models can be difficult to interpret and clinical feature relations remain incompletely characterized.

Abstract

from arXiv · show

This article reviews recent advances in applying natural language processing (NLP) to Electronic Health Records (EHRs) for computational phenotyping. NLP-based computational phenotyping has numerous applications including diagnosis categorization, novel phenotype discovery, clinical trial screening, pharmacogenomics, drug-drug interaction (DDI) and adverse drug event (ADE) detection, as well as genome-wide and phenome-wide association studies. Significant progress has been made in algorithm development and resource construction for computational phenotyping. Among the surveyed methods, well-designed keyword search and rule-based systems often achieve good performance. However, the construction of keyword and rule lists requires significant manual effort, which is difficult to scale. Supervised machine learning models have been favored because they are capable of acquiring both classification patterns and structures from data. Recently, deep learning and unsupervised learning have received growing attention, with the former favored for its performance and the latter for its ability to find novel phenotypes. Integrating heterogeneous data sources have become increasingly important and have shown promise in improving model performance. Often better performance is achieved by combining multiple modalities of information. Despite these many advances, challenges and opportunities remain for NLP-based computational phenotyping, including better model interpretability and generalizability, and proper characterization of feature relations in clinical narratives

1 INTRODUCTION

EHRs combine structured measurements and coded data with unstructured clinical narratives, creating resources for computational phenotyping. NLP expands phenotyping beyond structured-data approaches by mining clinically meaningful information from narratives.

  • EHRs contain structured data, such as diagnoses and laboratory values, alongside unstructured progress notes and discharge summaries.
  • Computational phenotyping automatically mines or predicts clinically significant phenotypes from structured data, narratives, or both.
  • Early studies commonly formulated phenotyping as supervised learning for constructing cohorts matching predefined phenotype criteria.
  • Structured-data studies use codes including ICD-9, ICD-10, SNOMED CT, RxNorm, and LOINC, while NLP historically focused mainly on term and keyword extraction.
  • EHR access is typically institutionally controlled, and database queries retrieve selected structured and unstructured data for research.

2 APPLICATIONS OF EHR-BASED COMPUTATIONAL PHENOTYPING

EHR-based computational phenotyping supports cohort identification, phenotype discovery, clinical trial screening, pharmacogenomics, safety analysis, and downstream genomic studies. These applications use structured records, clinical narratives, literature, and curated databases to identify clinically or scientifically relevant patterns.

  • Computational phenotyping supports diagnosis categorization, novel phenotype discovery, clinical trial screening, pharmacogenomics, DDI and ADE detection, and downstream genomics.
  • Diagnosis categorization enables automated identification of patient cohorts for secondary analysis across many diseases and disease subtypes.
  • Semi-supervised and unsupervised methods can identify traits from intrinsic data patterns with moderate or minimal expert guidance, supporting novel phenotype discovery.
  • Clinical-trial screening matches patient profiles extracted from structured data and narratives against eligibility criteria to reduce candidates for staff screening.
  • Phenotyping tools mine pharmacogenomics relations and detect drug interactions or adverse drug events from literature, clinical records, and medication data.
  • Computationally generated cohorts support GWAS and PheWAS by linking EHR phenotypes with genomic information and reducing labor-intensive annotation.

3 METHODS FOR NLP-BASED COMPUTATIONAL PHENOTYPING

NLP-based computational phenotyping spans keyword and rule-based systems, supervised and unsupervised learning, and deep learning. The review describes trade-offs among manual effort, labeling requirements, performance, phenotype discovery, generalizability, and interpretability.

  • Method overview: Computational phenotyping methods progressed from keyword searches and customized rules to supervised statistical learning and newer unsupervised approaches.
  • Keyword Search and Rule-based System: Keyword search uses terms, derivations, combinations, and regular expressions to extract phenotype indicators, but constructing keyword tables requires substantial manual effort.
  • Keyword Search and Rule-based System: 75% precision and 89% recall were achieved using 4,570 filtered keywords derived from 80 selected concepts.
  • Keyword Search and Rule-based System: Rule-based systems identify phenotypes by applying predefined rules to extracted concepts and contexts, with reported performance including a 92% overall kappa score and 99.6% document-level F-measure.
  • Supervised Statistical Machine Learning Algorithms: Supervised statistical methods learn classification patterns from labeled data and include logistic regression, Bayesian networks, SVMs, decision trees, and random forests.
  • Supervised Statistical Machine Learning Algorithms: SVMs repeatedly outperformed other learning algorithms in reviewed computational-phenotyping studies, with accuracies of 90% for smoking status, 87% for co-morbidity, and 82% for principal diagnoses in one study.
  • Unsupervised Learning: Unsupervised learning avoids expert annotation requirements and searches intrinsic data patterns, enabling clustering and discovery of phenotype subtypes.
  • Deep Learning: Deep learning models learn representations and intricate structures in high-dimensional EHR data, and reported studies found better performance than SVM, random forest, decision tree, or traditional rule-based systems.

4 MAKING NLP MORE EFFECTIVE

NLP-based computational phenotyping performance depends on model choice, data modality, feature extraction, and relation modeling. The reviewed studies favor carefully designed systems, heterogeneous data integration, and increasingly automated methods, while highlighting scalability and portability constraints.

  • 4.1 Comparison of Models: Model performance varies with data sources, features, training-set size, and target phenotype.
  • 4.1 Comparison of Models: Well-designed keyword and rule-based systems often perform well, but manually constructing their lists is laborious and difficult to scale.
  • 4.2 Combining Multiple Data Modalities: Combining structured EHR fields, clinical narratives, and external knowledge sources can improve prediction across computational phenotyping tasks.
  • 4.2 Combining Multiple Data Modalities: Narrative-based models and features appear less portable than those based on structured EHR fields, limiting generalizability across settings.
  • 4.3 Entity Recognition and Relation Extraction: Accurate entity recognition supplies concepts used as model features, while unsupervised methods support feature learning and novel relation discovery.
  • 4.3 Entity Recognition and Relation Extraction: Integrating entity detection with relation extraction can automate discovery and curation of biomedical knowledge, with unsupervised approaches offering adaptability across biomedical subdomains.

5 FUTURE WORK

Future NLP-based computational phenotyping research must address heterogeneous clinical narratives, limited model generalizability, poor interpretability, and complex feature relations. Promising directions include multimodal pipelines, models that learn phenotype phrases, and graph-based relation extraction.

  • 5.1 Information heterogeneity in clinical narratives: Clinical narratives vary across clinicians and often contain incomplete, ungrammatical, abbreviated, and context-limited documentation.External vocabulary modules can recognize variant expressions, but their performance remains controversial.
  • 5.2 Model generalizability: Rule-based systems remain difficult to generalize because adapting rules to new systems requires intensive human labor.High-dimensional feature spaces, data sparsity, and class imbalance also impede statistical and machine-learning approaches.
  • 5.2 Model generalizability: Combining multiple data sources and developing complete pipelines are proposed directions for improving computational phenotyping generalizability.These approaches address the limitations of adapting rules and handling sparse, high-dimensional, imbalanced data.
  • 5.3 Model interpretability: Convolutional neural networks may reduce annotation complexity by learning phenotype-associated phrases, but their interpretability remains unresolved.Interpretation of novel phenotypes discovered through unsupervised clustering is another major challenge.
  • 5.4 Characterizing the context of computational phenotyping: Generalized relation and event extraction can represent clinical conditions, mentions, and feature relations when relation structures are complex or not fully predefined.Graph methods are identified as a promising algorithmic direction for clinical-trial eligibility, diagnosis categorization, and pharmacogenomic networks.

6 CONCLUSION

The review finds that keyword, rule-based, and supervised machine-learning methods dominate NLP-based computational phenotyping, while unsupervised learning and multimodal integration are gaining attention. No single method dominates, and interpretability, generalizability, and clinical feature relations remain open challenges.

  • 6 CONCLUSION: Keyword search, rule-based systems, and supervised machine learning are the most widely used NLP methods for computational phenotyping.Well-designed keyword and rule-based systems often achieve high accuracy, while supervised classification is easy to train and test but requires labeled samples.
  • 6 CONCLUSION: Manual keyword and rule construction limits scalability and generalizability, whereas supervised methods require labor-intensive labeled training data.Performance can also vary with data sources, data types, and sample sizes.
  • 6 CONCLUSION: Unsupervised machine learning is gaining attention because it requires less human annotation and may discover novel phenotypes.The review also notes increasing interest in combining different data sources to improve model performance.
  • 6 CONCLUSION: Remaining challenges include improving model interpretability and generalizability and properly characterizing feature relations in clinical narratives.These issues continue to define research opportunities for NLP-based computational phenotyping.
Loading 1806.04820v2…