Source-linked AI summary
Natural Language Processing in Electronic Health Records in Relation to Healthcare Decision-making: A Systematic Review
Elias Hossain, Rajib Rana, Niall Higgins, Jeffrey Soar, Prabal Datta Barua, Anthony R. Pisani, Ph. D, Kathryn Turner}
TL;DR
Clinical NLP must organize and interpret increasingly large, varied EHR data while addressing information overload, privacy, and context-understanding challenges. This study reviews NLP applications and ML/DL approaches for EHRs, finding that several models suit unstructured free-text analysis while study assessment and data imbalance remain concerns.
Problem
Increasing EHR data volume, information overload, varied text formats, and privacy requirements complicate extracting clinically useful information.
Method
The study conducts a systematic review of clinical NLP applications, ML/DL architectures, and automated tools for analyzing EHR data.
Results
SVM, boosting techniques, LR, LSTM, RNN, and CNNs are appropriate for analyzing unstructured free-text data in downstream EHR applications.
Takeaways & Limitations
Clinical NLP research spans patient-risk analysis, summarization, named entity recognition, chart review, mental-health research, and medical language translation.
Takeaways & Limitations
The review excluded grey literature and did not assess publication bias, limiting its coverage of the available evidence.
Abstract
from arXiv · showhide
Background: Natural Language Processing (NLP) is widely used to extract clinical insights from Electronic Health Records (EHRs). However, the lack of annotated data, automated tools, and other challenges hinder the full utilisation of NLP for EHRs. Various Machine Learning (ML), Deep Learning (DL) and NLP techniques are studied and compared to understand the limitations and opportunities in this space comprehensively. Methodology: After screening 261 articles from 11 databases, we included 127 papers for full-text review covering seven categories of articles: 1) medical note classification, 2) clinical entity recognition, 3) text summarisation, 4) deep learning (DL) and transfer learning architecture, 5) information extraction, 6) Medical language translation and 7) other NLP applications. This study follows the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Result and Discussion: EHR was the most commonly used data type among the selected articles, and the datasets were primarily unstructured. Various ML and DL methods were used, with prediction or classification being the most common application of ML or DL. The most common use cases were: the International Classification of Diseases, Ninth Revision (ICD-9) classification, clinical note analysis, and named entity recognition (NER) for clinical descriptions and research on psychiatric disorders. Conclusion: We find that the adopted ML models were not adequately assessed. In addition, the data imbalance problem is quite important, yet we must find techniques to address this underlining problem. Future studies should address key limitations in studies, primarily identifying Lupus Nephritis, Suicide Attempts, perinatal self-harmed and ICD-9 classification.
1 introduction
EHR adoption has created large, heterogeneous data resources that require NLP and ML tools to convert clinical information into usable knowledge. This review surveys state-of-the-art approaches, applications, challenges, validation practices, and clinical implementations in EHR-focused NLP.
- 1 introduction: EHR adoption has rapidly expanded, producing large volumes of structured and unstructured healthcare data for provision, administration, and research.In the United States, adoption rose from 10% to nearly 96% between 2008 and 2017.
- 1 introduction: Human cognitive capacity is constrained when interpreting EHR big data, motivating computer-based tools that organize information, evaluate it, and recognize patterns.The review identifies data mining and NLP as essential components for converting extensive healthcare data into knowledge.
- 1 introduction: Clinical NLP remains challenging because clinicians use specialized language and idioms, while accurate model selection for clinical text analysis is still developing.Information overload, semantic and context understanding, and diverse text formats further complicate clinical NLP.
- 1 introduction: Earlier reviews often lacked PRISMA adherence, state-of-the-art model coverage, discussion of clinical tools and settings, and information about model validation or evaluation metrics.The authors position their review as addressing these gaps, including advanced methods such as transformer models.
- 1 introduction: The review examines ML and DL models, their feature extraction and evaluation approaches, EHR data types, preprocessing pipelines, study settings, applications, and automated tools.Its objectives also include identifying challenges in clinical-note categorization and examining how models are implemented for managing clinical notes.
2 Methodology
The review searched peer-reviewed literature from 2016 to 2022 across eleven databases using defined inclusion and exclusion criteria. The included studies primarily analyzed unstructured EHR free text with experimental designs and varied methods.
- 2 Methodology: The literature search covered eleven electronic databases and articles published from 2016 to 2022.The databases included major medical, engineering, scientific, and informatics outlets.
- 2 Methodology: Eligible studies used ML or DL models for clinical free-text classification, embeddings, summarisation, dialogue systems, concept representation, risk identification, entity recognition, or related EHR tasks.Only peer-reviewed journal articles or full conference papers were included.
- 2 Methodology: Preprints, preliminary or non-peer-reviewed work, editorials, and review papers were excluded, followed by quality examination of retrieved full texts.The exclusion criteria were applied after initial screening and before full-text analysis.
- 2 Methodology: 261 titles underwent initial title-and-abstract screening, while 101 additional records came from reference lists; 119 were excluded under the criteria and four duplicates were removed.The passage describes the search-output process represented in Figure 1.
- 2 Methodology: EHR was the most commonly used data type, most datasets were unstructured, and reviewed study designs were mainly experimental.Reported designs included 24 experimental studies, 9 cohort studies, and 1 case study.
3 Techniques used in the literature for Analysing EHR
The reviewed EHR-NLP techniques span patient risk analysis, advanced architectures, summarisation, information extraction, blockchain systems, chart review, and medical language translation. Studies apply diverse models to unstructured clinical text, while data access, annotation, and evaluation remain important constraints.
- EHR analysis techniques include patient risk prediction, advanced architectures, medical text summarisation, and other NLP applications.
- Patient Risk Analysis/Prediction: NLP-CAM and NLPmCAM use Confusion Assessment Method features to identify delirium from patient charts.The features include sudden onset and variable course, lack of attention, disordered thoughts, and altered awareness.
- Advanced Architectures to Analyse EHRs: BEHRT models disease progression, care delivery, and event timing to forecast future diagnoses from patients’ previous illnesses and symptoms.Compared with RETAIN, it showed an 8.0–13.2% increase in average accuracy for sickness-trajectory and illness-prediction tasks.
- Advanced Architectures to Analyse EHRs: MS-BERT was trained on more than 70,000 de-identified Multiple Sclerosis consultation notes and outperformed word2vec, CNN, and rule-based models in macro-F1.The model was evaluated by forecasting the Expanded Disability Status Scale.
- Advanced Architectures to Analyse EHRs: CheXbert uses BERT to learn radiology-report labels from rule-based annotations and expert comments, improving F1 scores on MIMIC-CXR.The approach addresses the limited availability and high production cost of human annotations.
- Medical Text Summarising: BT-45 generates textual summaries from approximately forty-five minutes of neonatal intensive-care information through signal analysis, interpretation, planning, and realisation.Human evaluators found the summaries inefficient because disparate data were difficult to integrate.
- Medical Text Summarising: Remote-supervision systems use medical BERTs and diagnostic-code groupings to select short EHR phrases, significantly outperforming unsupervised baselines.The authors proposed testing whether limited direct supervision could further improve performance.
- Other NLP Applications: Blockchain-based healthcare work used Hyperledger Fabric alongside NLP to support communication and record tracking.The framework includes participants, system elements, and transaction procedures; prescription images were preprocessed with grayscale conversion, Otsu thresholding, erosion, and dilation.
4 Analysis of the literature
The review analyses how retrieved studies use machine learning and deep learning to process EHRs, including their data, preprocessing, models, evaluation, features, and clinical settings. The supplied passages identify the review’s planned analysis structure but do not report substantive findings for this section.
- The analysis covers data types and quantities, clinical free-text preprocessing, frequently used ML and DL models, model comparisons, evaluation matrices, feature extraction, and clinical settings.
4.1 Data Type and Quantity
The reviewed studies predominantly used EHR data, especially unstructured clinical text, but restricted data availability and insufficient preprocessing detail limited reproducibility.
- Clinical notes were the most frequently recorded EHR information type, appearing in 14 studies.
- Most EHR data were not publicly accessible because patient confidentiality constrained data availability.
- Most reviewed studies used electronic health record data, with data formats reported as primarily unstructured.
- The reviewed studies included observational studies (n = 8), experimental studies (n = 24), and case studies (n = 1).
- Researchers often failed to explain details of clinical-text preprocessing methods.
4.2 Models
The review identified diverse machine-learning and deep-learning models for clinical NLP, with classification and prediction as the most prominent tasks and neural networks the most frequent deep-learning model.
- Reviewed studies applied Logistic Regression, Support Vector Machine, XGBoost, AdaBoost, Random Forest, Naïve Bayes, Gradient Boosting, and Decision Tree models.
- Support Vector Machine and boosting algorithms were the most widely used classical machine-learning models for EHR data.
- Support Vector Machine use for clinical free-text analysis increased by 95% in recent years, while Decision Tree use was lowest at 40%.
- Evaluation metrics were not discussed in many deep-learning studies, and the review could not determine why.
- Neural Network was the most frequently used deep-learning model for analysing clinical free text in EHRs.
- Other identified models included LSTM, BI-LSTM, CNN, ResNet, transfer learning, RNN, GRU, and representation learning.
4.3 Comparison of frequently utilised models
The review compares commonly used machine-learning and deep-learning models in clinical free-text processing, focusing on their general advantages and disadvantages.
- The comparison is intended to help readers understand the core information of machine-learning and deep-learning models in clinical free-text contexts.
- Table 5 compares popular machine-learning and deep-learning models applied to electronic health records.
4.4 Model Evaluation Metrics
The reviewed studies primarily used AUC, Accuracy, Precision, Recall, and F1-score to evaluate model performance, with equations provided for precision, recall, and F1-score.
- AUC, Accuracy, Precision, Recall, and F1-score were the primary evaluation metrics across the reviewed articles.
- AUC differentiates dataset classes, with higher AUC indicating better separation of positive and negative classes.
- Equations (1)–(4) were applied primarily to determine precision, recall, and F1-score.
- Accuracy is calculated as TP + TN divided by TP + TN + FP + FN.
4.5 Word Embedding/Feature Extraction Methods
The reviewed EHR studies used both traditional feature extraction and word-embedding approaches, with TF-IDF, BOW, and GloVe appearing most frequently.
- 4.5 Word Embedding/Feature Extraction Methods: Feature extraction approaches for EHR narratives included word weighting, word embedding, and open-source tools.Examples included TF-IDF, BOW, Word2vec, GloVe, and FastText.
- 4.5 Word Embedding/Feature Extraction Methods: TF-IDF was used in 7 selected papers, while BOW and GloVe were each used in 5.
- 4.5 Word Embedding/Feature Extraction Methods: TF-IDF and BOW served as word-weighting techniques, whereas Word2vec and GloVe were favored for word embeddings.
4.6 Automated tools
Automated ML tools support healthcare applications through commercial and open-source platforms, but their availability and capabilities differ by data type, coding requirements, and cost.
- 4.6 Automated tools: Commercial AutoML solutions from Google, Amazon, Microsoft, and JADBIO are chargeable and require no coding.
- 4.6 Automated tools: Free AutoML technologies require minimal local-environment coding and have some limitations compared with fee-based solutions.
- 4.6 Automated tools: Existing AutoML technologies primarily target well-defined unstructured data and are not commercially available for structured data.
- 4.6 Automated tools: Most unstructured data was processed with commercial solutions, whereas structured data was often handled by open-source tools and clinical AutoML platforms.
- 4.6 Automated tools: Commercial platforms such as Amazon Recognizer, Apple CreateML, Microsoft AutoML, and Google AutoML generally required no coding and were relatively user-friendly.
5 Research Viewpoint
The review identifies information overload, semantic and contextual ambiguity, varied clinical text formats, privacy constraints, data shortages, and inadequate model assessment as major clinical NLP challenges. It also highlights domain-specific priorities and future directions, including better annotation, longitudinal data, privacy standards, and imbalance mitigation.
- Core Challenges in Clinical NLP: Information overload makes it difficult to locate significant information in vast clinical datasets, while semantic and contextual understanding remain challenging for summarisation systems.Clinical NLP systems must also handle diverse text formats from multiple sources.
- Spelling Correction and Negation Detection: Clinical abbreviations, misspellings, and negation can alter or obscure clinical meaning during text analysis.NegEx detects negation but can fail in complex phrases because it disregards contextual relationships between words.
- Lack of Medical Data: Data shortages constrain medical NLP because supervised models require sufficient training data, while privacy and liability concerns limit health-system data sharing.Reviewed studies primarily obtained data directly from hospitals, and scarcity limited use of transformer-based and advanced deep-learning methods.
- Sensitivity of Medical Data and Privacy: Sanitisation protects privacy but can compromise data value and integrity, may not ensure adequate sharing, and can leave re-identification risks.Clinical NLP research must also address informed consent and sensitive-data protection under HIPAA and GDPR.
- Model Assessment and Point of View of Adopted Models: Existing studies often lacked model validation indicators, leaving generalisation and overfitting insufficiently assessed.The review specifically identifies missing validation practices such as K-Fold Cross Validation.
- Model Assessment and Point of View of Adopted Models: Clinical NLP research commonly used traditional TF-IDF and BOW features, while advanced word embeddings were comparatively uncommon and semantic information remained difficult to obtain.The review identifies FastText, GloVe, and BERT as methods that can mitigate this difficulty.
- Future Research Directions: Future work should improve goal-of-care annotation, ontology development, and bias reduction, while suicide prediction requires better predictors and large longitudinal datasets.The review also calls for stronger privacy standards for federated learning and secure multi-party computing.
6 Conclusion
The review surveys NLP applications in EHRs, finding several established approaches for unstructured clinical text while identifying continuing challenges for deep learning and transformers.
- The review covers EHR-NLP applications including risk analysis, medical summarisation, clinical entity recognition, mental health research, chart review, negation identification, and translation.
- Dementia and geriatric mental health are identified as promising research areas involving ongoing exploration of task-suited feature extraction methods.
- SVM, boosting, logistic regression, LSTM, RNN, and CNN methods are considered appropriate for analysing unstructured free text in downstream EHR applications.
- Deep learning remains difficult in biomedical NLP because of limited data availability, domain-specific textual complexity, interpretability concerns, and substantial data and GPU requirements.
- Transformer-based NLP methods for free-text analysis are not yet widely used, with conventional approaches still preferred.
Funding Statement
The study received no specific grant funding.
- The study received no specific grant from any funding agency.
- No specific funding was received from public-sector agencies.
- No specific funding was received from commercial or not-for-profit agencies.