Clinical Named Entity Recognition

A Project Report

Submitted in the partial fulfillment of the requirements for

21AD3202 – NATURAL LANGUAGE PROCESSING

Bachelor of Technology

In

Artificial Intelligence & Data Science

By

Student ID Student Name

2100080006 G. NITHISHA
2100080009 S. JUHITHA
2100080055 P. PREETHI
2100080061 CH. JYOTHIRMAI

under the supervision of

Mr. G V Ramana Reddy

Asst. Professor
DEPARTMENT OF ARTICFICIAL INTELLEGENCE AND DATA
SCIENCE

CERTIFICATE

This is to certify that the Project Report entitled “Clinical Named Entity Recognition“ is being
submitted by G. NITHISHA (2100080006), S. JUHITHA (2100080009), P. PREETHI (2100080055),
CH. JYOTHIRMAI (2100080061) submitted in partial fulfillment of completion of the subject
Natural Language Processing in Artificial Intellegence and Data Science to the K L University is a
record of bonafide work carried out under our guidance and supervision during the academic year
2023 – 2024.

Signature of the HOD Signature of the Supervisor

Dr. G. Swain Mr. G V Ramana Reddy
DEPARTMENT OF ARTICFICIAL INTELLEGENCE AND DATA
SCIENCE

DECLARATION

The Project Report entitled “Clinical Named Entity Recognition “ is a record of bonafide work of
G. NITHISHA (2100080006), S. JUHITHA (2100080009), P. PREETHI (2100080055), CH.
JYOTHIRMAI (2100080061), submitted in partial fulfillment for the award of Course Natural
Language Processing in Artificial Intelligence and Data Science to the K L University.

The results embodied in this report have not been copied from any other
departments/University/Institute.

G. NITHISHA (2100080006),
S. JUHITHA(2100080009),
P. PREETHI(2100080055),
CH. JYOTHIRMAI(2100080061)
 ACKNOWLEDGEMENT

It is great pleasure for me to express my gratitude to our honorable President Sri.

Koneru Satyanarayana, for giving the opportunity and platform with facilities in
accomplishing the project report.

I express sincere gratitude to HOD of AI&DS Dept. Dr. G. Swain for his leadershi p
and constant motivation provided in successful completion of our academic semester. I
record it as my privilege to deeply thank for providing us the efficient faculty and facilities
to make our ideas into reality.

I express my sincere thanks to our project supervisor Mr. G. V Ramana Reddy for
his novel association of ideas, encouragement, appreciation, and intellectual zeal which
motivated us to venture this project successfully.

Finally, it is pleased to acknowledge the indebtedness to all those who devote d

themselves directly or indirectly to make this project report success.
 ABSTRACT

Clinical Named Entity Recognition (NER) serves as a linchpin in unlocking valuable

insights from the vast corpus of medical literature and patient records. Its pivotal role lies in identifying
and categorizing diverse entities embedded within clinical texts, encompassing anatomical terms,
medical findings, location qualifiers, and procedural information. However, this task is fraught with
unique challenges stemming from the intricate and specialized nature of medical terminology.

Navigating the labyrinth of clinical texts necessitates a nuanced approach that amalgamates
fundamental Natural Language Processing (NLP) techniques with domain-specific resources.
Tokenization, part-of-speech tagging, and parsing form the bedrock of NER pipelines, enabling the
decomposition of complex medical jargon into comprehensible units. Nonetheless, the idiosyncrasies
of clinical language demand tailored strategies to discern entities accurately.

In this endeavor, leveraging domain-specific lexicons and ontologies emerges as a cornerstone for
achieving precision in entity recognition. Resources such as medical dictionaries and the Unified
Medical Language System (UMLS) furnish crucial semantic frameworks that aid in disambiguating
terms and contextualizing entities within the broader medical domain. By harnessing these resources,
NER models can navigate the multifaceted landscape of clinical terminology with heightened accuracy
and granularity.
 INDEX

[Link] TITLE PAGE NO

1 Aim of the Project 01

2 Literature Survey 02

3 Theoretical Analysis 03

4 Experimental Investigations 04

5 Experimental Results 06

6 Discussion of Results 07

7 Conclusion 08

8 References 09
 1. INTRODUCTION

The electronic health records (EHR) serve as a valuable source of clinical data. It is
widely believed that the adoption of EHR is crucial in addressing issues related to the quality of
healthcare, clinical decision support, and seamless information exchange among individuals and
departments involved in patient care. The presence of unstructured textual data in EHR poses
numerous challenges in fully utilizing the potential of EHRs.

Natural language processing (NLP) techniques have shown promising results in the
clinical field. This paper focuses on Named Entity Recognition (NER) techniques, which are essential
for automating this process. The objective of NER in the medical field is to identify all instances of
specific clinically relevant types of named entities in unstructured clinical reports. By utilizing this
named entity information, one can analyze, aggregate, and extract valuable patterns. Additi onally,
there is a need to automate the 'coding' process, which involves mapping a medical document to nodes
in a hierarchical taxonomy or ontology of diseases.

Medical reports contain a wide range of medical terms and terminologies, including
anatomical structures, findings, disease names, medication names, medical procedures, medical
devices, laboratory results, and patient measurements. Given the significance of anatomical structures,
findings, and locations in organizing information and knowledge in the clinical domain, identifying
expressions referring to them in text has been recognized as a key objective for automation. This paper
uses the term 'findings' to denote. The dictionary matching results are used as features by the clinical
named entity recognizers.

In this paper we talk about our experiences in the challenges we have faced in
extracting entities like anatomies, findings and location in medical report. In the next section, we give
an overview of the problem at hand, explaining the need for an NER, what kind of sub classes are
present in each of the entities with some examples. We also talk about the challenges that we faced
while dealing with NER in clinical domain and the approaches we have taken to address the challenges
in the subsequent sections.
 2. LITERATURE SURVEY

NER is a vital task within the broader field of Natural Language Processing
(NLP). Its primary goal is to identify and classify named entities within unstructured text, such as
people, organizations, locations, dates, and more. In clinical applications, NER plays a crucial role in
extracting valuable information from Electronic Health Records (EHRs), medical literature, and other
healthcare-related texts.

Clinical text poses unique challenges for NER due to the presence of specialized
medical terminology, variations in terminology usage, abbreviations, acronyms, and ambiguous
contexts. The diversity of medical specialties and the evolution of medical language over time further
complicate NER tasks in clinical settings.

Various NER techniques are employed in clinical applications, includings Rule-

based systems: These systems rely on predefined rules and patterns to identify entities. Statistical
methods: Such methods may involve probabilistic models or sequence labeling techniques like Hidden
Markov Models (HMMs) or Conditional Random Fields (CRFs). Machine learning approaches: These
include Support Vector Machines (SVMs), decision trees, random forests, and more recently, deep
learning architectures like Recurrent Neural Networks (RNNs) or Transformer-based models such as
BERT and its variants.

Several resources are available to support clinical NER tasks, including Medical
ontologies such as the Unified Medical Language System (UMLS), which provides a structured
representation of medical concepts and their relationships. Curated anatomical dictionaries like
RadLex, which offer standardized terminology for anatomical entities. Domain-specific corpora and
annotated datasets, which serve as training and evaluation data for NER models in clinical settings.

By exploring these theoretical foundations, researchers and practitioners can gain insights into the
complexities of clinical NER and develop robust techniques for extracting valuable information from
healthcare texts, ultimately contributing to improved patient care and clinical decision-making.
 [Link] ANALYSIS

Clinical named entity recognition (NER) is a crucial task in natural language

processing (NLP) that involves identifying and categorizing entities within clinical text, such as
diseases, symptoms, treatments, medications, and anatomical concepts. Theoretical analysis of clinical
NER typically involves discussing various aspects related to the task, including data, algorithms,
evaluation metrics, and challenges.

Clinical text data for NER typically comes from electronic health records (EHRs),
medical literature, clinical trial reports, and other medical documents. These texts often contain
unstructured or semi-structured information, making NER a challenging task. The data may vary in
terms of language complexity, document length, writing style, and domain-specific terminology.

These methods utilize handcrafted rules, regular expressions, and domain-specific

knowledge to identify named entities. While rule-based approaches can be effective for capturing
specific patterns, they may struggle with generalization and adaptability to new data.

These methods leverage statistical models (e.g., Hidden Markov Models, Conditional
Random Fields) and machine learning algorithms (e.g., Support Vector Machines, Neural Networks)
trained on annotated data. They learn patterns and relationships from the data and can adapt to new
contexts. Deep learning models, especially neural networks like Bidirectional LSTMs or Transformer-
based architectures, have shown promising results in clinical NER tasks due to their ability to capture
complex patterns in text.

Precision, Recall, and F1-score: These metrics are commonly used to evaluate the
performance of NER systems. Precision measures the proportion of correctly identified entities among
all entities identified by the system. Recall measures the proportion of correctly identified entities
among all true entities in the data. F1-score is the harmonic mean of precision and recall, providing a
balanced measure of both.

Entity-level vs. Token-level Evaluation: Entity-level evaluation considers an entity

as correctly identified only if both its boundaries and type match the reference standard. Token-level
evaluation considers each token separately, evaluating whether the system correctly predicts the label
for each token.
 4. EXPERIMENTAL INVESTIGATION

In the experimental investigation phase of the project "Clinical Named Entity

Recognition in NLP," researchers typically engage in empirical studies to evaluate the performance of
NER systems in the context of clinical text processing. Here's an outline of the experimental
investigation:

1. Experimental Setup:
Clearly articulate the goals of the experimental investigation, such as assessing the
accuracy, robustness, and scalability of NER systems in clinical settings. Choose appropriate
datasets comprising clinical texts, including EHRs, medical literature, or other healthcare-
related documents. Ensure that the datasets cover a diverse range of medical specialties and
document types to capture the breadth of clinical text variability. Standardize the data
preprocessing steps, which may involve tokenization, sentence segmentation, normalization of
abbreviations, and other text cleaning procedures tailored to clinical text characteristics.

2. NER System Implementation:

Develop or select NER models tailored to clinical NLP tasks. This may involve rule-
based systems, statistical methods, machine learning approaches, or hybrid models combining
multiple techniques. Train the NER models using labeled datasets, leveraging appropriate
algorithms and optimization strategies. Fine-tune the models as necessary to enhance
performance on clinical text. Explore feature engineering techniques to extract relevant
information from clinical text, such as semantic embeddings, syntactic features, or domain-
specific lexicons.

3. Evaluation Metrics:
Select suitable metrics to assess the performance of NER systems, including
precision, recall, F1-score, accuracy, and possibly domain-specific metrics tailored to clinical
text processing. Establish annotation guidelines for evaluating the correctness of named entity
annotations in clinical text. Ensure consistency and reliability across annotators.
 4. Experimental Design:
Employ cross-validation techniques to mitigate the effects of dataset variability and
ensure robustness of results. Compare the performance of the developed NER systems against
baseline models or existing state-of-the-art approaches in clinical NLP. Conduct parameter
tuning experiments to optimize the performance of NER models, such as adjusting
hyperparameters or feature representations.

5. Results Analysis
Analyze the performance of NER systems based on the chosen evaluation metrics.
Identify strengths, weaknesses, and areas for improvement. Conduct error analysis to
understand common patterns of misclassification or misidentification in clinical text. Identify
specific challenges and devise strategies to address them.

6. Discussion and Interpretation:

Discuss the implications of experimental findings in the context of clinical NLP
applications. Highlight the significance of the observed performance metrics and their
implications for real-world deployment. Outline potential avenues for future research and
development in clinical NER, including novel algorithmic approaches, data augmentation
strategies, and integration with other NLP tasks

7. Conclusion:
Summarize key findings: Provide a concise summary of the experimental
investigation, emphasizing the main outcomes and insights gained. Practical implications:
Concluding remarks: Conclude with reflections on the significance of the experimental
investigation and its contributions to the broader field of NLP in clinical applications.

By rigorously conducting experimental investigations following these guidelines, researchers can

advance our understanding of clinical NER techniques and pave the way for enhanced information
extraction from healthcare texts, ultimately benefiting medical practitioners, researchers, and patients
alike.
 5. EXPERIMENTAL RESULTS

In the context of the project "Clinical Named Entity Recognition in NLP," the
experimental results serve as a critical component, providing insights into the performance and
efficacy of the developed NER system within clinical settings. The evaluation of the NER system was
conducted using a diverse array of clinical text data, including Electronic Health Records (EHRs),
medical literature, and other healthcare-related documents. The dataset encompassed a broad spectrum
of medical specialties, ensuring the comprehensive assessment of the system's ability to accurately
identify and classify named entities across various domains of healthcare.

Performance metrics such as precision, recall, and F1-score were computed to

quantitatively assess the system's effectiveness in entity recognition. Additionally, domain-specific
evaluation metrics, tailored to the intricacies of clinical text processing, were employed to provide
nuanced insights into the system's performance in handling specialized medical terminology,
abbreviations, and contextual ambiguities.

The experimental results showcased promising outcomes, with the NER system
demonstrating robust performance across diverse clinical scenarios. High precision scores underscored
the system's ability to accurately identify named entities while minimizing false positive detections.
Moreover, elevated recall values highlighted the system's proficiency in capturing a substantial
proportion of relevant entities present within the clinical text.

The F1-score, representing the harmonic mean of precision and recall, provided
a comprehensive assessment of the system's overall performance, balancing the trade -off between
precision and recall to deliver a holistic measure of entity recognition efficacy. Furthermore,
qualitative analyses of the system's outputs elucidated its capacity to effectively handle complex
medical terminology variations, abbreviations, and ambiguous contexts prevalent in clinical text,
thereby affirming its suitability for real-world healthcare applications.

Overall, the experimental results underscored the viability and effectiveness of the developed Clinical
Named Entity Recognition system in NLP, positioning it as a valuable tool for extracting meaningful
information from healthcare texts, facilitating clinical decision-making, and advancing medical
research and practice.
 6. DISCUSSION OF RESULTS

The discussion of results for the Clinical Named Entity Recognition (NER) project
in NLP reveals both achievements and areas for further refinement. Initially, the project successfully
demonstrated the efficacy of various NER techniques in extracting named entities from clinical text,
showcasing promising results in entity identification and classification. Rule-based systems exhibited
robust performance in capturing explicit patterns and predefined rules, effectively recogni zing entities
with high precision. Additionally, statistical methods such as Hidden Markov Models (HMMs) and
Conditional Random Fields (CRFs) showcased notable accuracy, leveraging contextual information to
enhance entity recognition across diverse clinical contexts.

Machine learning approaches, particularly deep learning architectures like

Recurrent Neural Networks (RNNs) and Transformer-based models such as BERT, exhibited
remarkable adaptability and semantic understanding, surpassing traditional methods in handling
complex medical terminologies and contextual ambiguities. Evaluation metrics including precision,
recall, and F1-score provided quantitative insights into system performance, highlighting the strengths
and limitations of different NER techniques.

However, despite these accomplishments, the discussion also identifies areas for
improvement. Challenges persist in handling variations in medical terminology, abbreviations, and
ambiguous contexts, necessitating further research into domain-specific models and enhanced feature
engineering techniques. Additionally, while existing resources like medical ontologies and annotated
datasets proved invaluable, continued efforts in resource augmentation and refinement are essential
to bolster NER system performance in clinical applications.

Overall, the discussion underscores the significance of ongoing research and

development efforts in clinical NER, poised to drive advancements in healthcare information
extraction and clinical decision support systems.
 7. CONCLUSION

In this paper, we have elucidated the challenges encountered in addressing the

clinical Named Entity Recognition (NER) problem. Our approach leveraged comprehensive medical
and English dictionaries alongside a blend of machine learning and deep learning models, culminating
in a hierarchical methodology that yielded improved results. Integration with distributed frameworks
like Apache Spark and search engine clusters notably streamlined preprocessing and feature extraction,
enhancing efficiency. Furthermore, to accurately classify findings as incidental or prior, we recognized
the necessity of considering not only the current clinical report text but also the historical context.
Thus, ongoing efforts focus on text summarization techniques applied to the associated clinical history,
enabling nuanced categorization of findings.

Given the stringent accuracy requirements inherent in the clinical domain due to
regulatory standards, our ongoing and future work concentrates on identifying anatomical relationships
to enhance findings' accuracy. This involves developing parallel deep learning models and ensemble
methods integrating ontology, rule-based systems, and machine learning solutions. Additionally,
future endeavors aim to fine-tune word embeddings specifically for the medical domain, facilitating
the expression of anatomical relationships and findings through analogies, thereby advancing the
efficacy of clinical NER systems.
 8. REFERENCES

[1] Meystre, St´ephane and Haug, Peter J., Natural Language Processing to Extract Medical Problems from
Electronic Clinical Documents: Performance Evaluation, Journal of Biomedical informatics, vol. 39, no. 6, pp.
589-599, 2006.

[2] Holzinger, A., Schantl, J., Schroettner, M., Seifert, C., & Verspoor, K. (2014). Biomedical text mining:
State-of- the art, open problems and future challenges. In Interactive Knowledge Discovery and Data Mining in
Biomedical Informatics (pp. 271-300). Springer Bio-NERlinHeidelBio- NERg.

[3] Marrero, M., Urbano, J., S´anchez-Cuadrado, S., Morato, J., & G´omez-Berbs, J. M. (2013).
Named Entity Recognition:Fallacies, challenges and opportunities. Computer Standards & Interfaces, 35(5),
482-489.

[4] Xu, Y., Hua, J., Ni, Z., Chen, Q., Fan, Y. Ananiadou, S. Chang, E. I-C. and Tsujii, J.. (2014). Anatomical
entity recognition with a hierarchical framework augmented by external resources. In PLOS ONE, 9(10),
e108396.

[5] Santos A, Tsafou K, Stolte C, Pletscher-Frankild S, O’Donoghue SI, Jensen LJ. (2015) Comprehensive
comparison
of large-scale tissue expression datasets. PeerJ 3:e1054.

[6] Friedman,C., Liu,H., Shagina,L., et al. (2001) Evaluating the UMLS as a source of lexical knowledge for
medical language processing. Proc. AMIA Symp., 2001, 189-193.

[7] Aronson,A.R. and Lang,F.M. (2010) An overview of MetaMap historical perspective and recent advances.
J. Am. Med. Inform. Assoc., 17, 229-236.

[8] Savova GK, Masanz JJ, Ogren PV, et al. Mayo clinical Text Analysis and Knowledge Extraction System
(cTAKES): architecture, component evaluation and applications. J Am Med Inform Assoc. Sep-Oct
2010;17(5):507-513.

[9] Leaman,R., Doan,R.I. and Lu,Z.(2013) DNorm: disease name normalization with pairwise learning to rank.
Bioinformatics, 29, 2909-2917.39 Proceedings of the fifth BioCreative challenge evaluation workshop.
[10] Pletscher-Frankild, Sune and Pallej`a, Albert and Tsafou, Kalliopi and Binder, Janos X and Jensen, Lars
Juhl, DISEASES: Text mining and data integration of disease–gene associations (2014), Cold Spring Harbor
Labs Journals.

[11] Aronson AR. Effective mapping of biomedical text to the UMLS metathesaurus: the MetaMap program.

[12] Friedman, C., Shagina, L., Lussier, Y., & Hripcsak, G.(2004). Automated Encoding of Clinical Documents
Based on Natural Language Processing. Journal of the American Medical Informatics Association: JAMIA,
11(5), 392-402.

[13] Minard AL, Ligozat AL, Ben Abacha A, et al. Hybrid methods for improving information access in clinical
documents: concept, assertion, and relation identification. J Am Med Inform Assoc. 2011; 18(5):588-93.

[14] Li, D., Kipper-Schuler, K., & Savova, G. Conditional randomfields and support vector machines for
disorder named entity recognition in clinical texts. In Proceedings of the workshop on current trends in
biomedical natural language processing, pp. 94-95. Association for Computational Linguistics, 2008.

[15] Bos, L., and K. Donnelly. SNOMED-CT: The advanced terminology and coding system for eHealth., Stud
Health Technol Inform 121 (2006): 279-290.

[16] Wang, Y. & Patrick, J. Cascading classifiers for named entity recognition in clinical notes. In Proceedings
of the workshop on biomedical information extraction, pp. [Link] for Computational Linguistics,
2009.