A modern framework for assessing patient histories and conducting clinical research has been developed as the number of clinical narratives evolves. To discover the knowledge from such clinical narratives, clinical entity recognition and relation extraction tasks were performed subsequently in existing approaches, which resulted in error propagation. Therefore, a novel end-to-end clinical knowledge discovery strategy has been proposed in this paper. The clinical XLNet was used as a base model for handling the discrepancy issue. To predict the dependent clinical relation association, the multinomial Naïve Bayes probability function has been incorporated. In order to improve the performance of the proposed strategy, it takes into account entity pairs presented consecutively through the multi-head attention layer. Tests have been conducted using the N2C2 corpus, and the proposed methodology achieves a greater than 20% improvement in accuracy over existing neural network-based and transformer-based methods.
Research Article
Clinical XLNet-based End-to-End Knowledge Discovery on Clinical Text Data using Natural Language Processing
https://doi.org/10.21203/rs.3.rs-2069629/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this older preprint version
Natural language processing
Electronic health record
Knowledge discovery
Entity recognition
Relation extraction
Clinical XLNet
The growing use of electronic health record systems (EHRs) in health care generates a vast growth of actual data, presenting new clinical investigation prospects [1]. EHR records include demographic information about patients, progress notes, problems, medicines, vital signs, prior medical histories, vaccinations, laboratory data, and radiological results. The electronic health record is a set of interconnected patient data systems [2-3]. The EHR includes the infrastructure that links these networks, as well as databases, interfaces, physician order entry, electronic communication systems, and clinical workstations. EHR systems are intended to improve the quality of patient care [4-5] with reduced utilizing resources, producing a data stream for computerized billing, and also providing electronic assistance for secondary users.
As therapeutic reports contain a significant quantity of important clinical information, natural language processing (NLP) strategies of artificial intelligence has been applied to extract information from these narratives. A natural language processing system employs computers to communicate with humans using their natural languages (not computer languages) [6-7]. In research on respiratory illness, as well as in other clinical areas, artificial intelligence has been demonstrated to be capable of recognizing and interpreting pulmonary function tests, detecting asymptomatic left ventricular dysfunction from a 12-lead EKG, and identifying early breast cancer stages from a biopsy [8-10]. There are a number of potential benefits that EHR-based informatics studies could offer beyond those explored in these studies. A massive quantity of complete continuous medical information is acquired, maintained, also made accessible digitally, encompassing clinical background, blood tests, drugs, therapies, and therapeutic intervention plans.
Clinical Named Entity Recognition (NER) is one of the vital steps in extracting patient data from medical records. Clinical NER has gained several research attention because it is a necessary step in the clinical data mining process. In several aspects, medical named entity recognition differs from generic NER. The massive quantity of alternative spellings and synonyms causes an expansion of vocabulary. This lowers the performance of medicine interpretation. The Clinical NER overcomes the challenges faced by this NER [11-13]. The primary goal of Clinical NER is to recognize and categorize clinical words in clinical data, such as symptoms, drugs, and therapy entity boundaries, and category labels are often anticipated simultaneously when approaching the topic as a sequence labeling issue [14]. Many studies have been undertaken to extract named elements from clinical literature through machine learning methodologies and deep neural network-based approaches [15-16]. The existing conventional NER models all use a neural network design that is devoid of handcrafted features, making them more flexible to various applications, languages, and domains [17-21].
One such task in knowledge discovery is relation extraction, which tries to discover and describe the semantic correlation between biological/clinical elements. The sorts of relationships might vary based on the genres and domains, including gene-protein interactions, medication interactions, or medical ideas (problems, treatments, or tests). The research community has conducted much competitive devaluation in connection extraction for biomedical texts in recent decades, resulting in a rise in attention. Shared task systems currently in use [22] use guided approaches and have a range of traits [23-25]. Every investigation however has one problem in widely accepted: the influence of various sorts of attributes has not been properly investigated, and intra-relationships between sentences have not yet been discovered. To address the aforementioned issues, a fresh information retrieval methodology must be devised.
The main contribution of this research is to excerpt knowledge from clinical text data using natural language processing and relation extraction.
- An end-to-end clinical knowledge discovery has been introduced based on the clinical XLNet model, whereas entity recognition, as well as relation extraction, are performed solely.
- Both the independent and dependent clinical relation associations are obtained with joint conditional and multinomial naïve Bayes probability functions.
- The relation between entity pairs presented in the same sentence as well as the consecutive sentences are obtained to discover the great knowledge from clinical narratives.
This paper has been structured as follows: Section 2 provides a review of recent literature, section 3 discusses the proposed methodology, section 4 discusses the results of the implementation, and section 5 concludes the paper.
This section provides a survey of some existing research related to clinical knowledge discovery and also identifies the gaps presented in those research.
Chen, et al [26] developed a framework for knowledge-enhanced reading comprehension (KRC) to extract biomedical relationships by leveraging prior knowledge and reading comprehension. For each relation, questions are created, turning the relation extraction process into a question-and-answer exercise. Additionally, an attention interaction mechanism with knowledge-enhanced representation was developed to aid in biomedical relation extraction by integrating the knowledge representation based on the RC framework.
Alimova, et al [27] presented a machine learning approach for clinical relation extraction with the knowledge base and BioSentVec embedding features. They have evaluated four distinctive characteristics of drug-related relation extraction such as distance-based, word-based, and knowledge also embedding. Carry out investigations with two standard corpora of medical records from MADE 2018 and n2c2 2018 shared tasks to comprehensively examine the influence of those characteristics with traditional distance- and word-based features.
Hong, et al [28], offer the BERE machine learning approach for extracting biomedical relationships from large-scale narrative collections automatically. BERE leverages a network for aggregating features to produce predictions after examining all relevant statements and a hybrid encoding network that better represents every phrase both semantically and syntactically. Furthermore, BERE can be trained using a strategy for remote monitoring without the need for any annotations made by humans. BERE has shown good results in extracting biomedical relationships, in addition to being able to find significant relationships, through comprehensive testing.
Alodadi, et al [29] proposed a method for achieving the use of evidence-based medicine by bridging knowledge discovery in clinical notes and the vast biomedical literature repositories. They have utilized natural language processing, text mining as well as data mining methods to excerpt, analyze and discover the knowledge from the clinical text data. Whereas, the clinical notes are extracted through NLP and then mapped to the UMLS, which might be utilized for identifying the various mentions of entities and to estimate the entity weights. The relation extraction is carried out by a rule-based mining algorithm.
Piad-Morffis, et al [30] propose an annotation model for natural language text that captures a considerable percentage of its semantics. The annotation model's structure is shown, together with samples of annotated phrases and a brief discussion of each semantic function and relation described. This study concentrates on the utilization of clinical texts in Spanish. The suggested annotation model, however, could be broadened to other domains and languages. A sample of annotated sentences, recommendations, and appropriate annotation tool configuration files are also given.
Table 1
Summary of Literature Survey
|
Sl. No |
Authors and Title |
Year |
Remarks |
|
|
Chen, et al Biomedical relation extraction via knowledge-enhanced reading comprehension |
2022 |
Developed a framework for knowledge-enhanced reading comprehension (KRC) to leverage prior knowledge and reading comprehension for extracting biomedical relationships. |
|
|
Alimova et al Multiple features for clinical relation extraction: a machine learning approach. |
2020 |
Presented machine learning approach for clinical relation extraction with knowledge and sentence embedding features. |
|
|
Hong et al A novel machine learning framework for automated biomedical relation extraction from large-scale literature repositories |
2020 |
BERE machine learning approach for extracting biomedical relationships from large-scale narrative collections automatically |
|
|
Alodadi et al Linking Knowledge Discovery In Clinical Notes And Massive Biomedical Literature Repositories |
2019 |
Clinical notes are extracted through NLP and mapped to UMLS. The relation extraction is carried out by a rule-based mining algorithm |
|
|
Paid-Morffis et al A general-purpose annotation model for knowledge discovery: A case study in Spanish clinical text |
2019 |
Proposed an annotation model that captures semantics from natural language text. The study concentrated on Spanish clinical text, however, could be broadened to other domains and languages. |
Table 1 shows the summary of the literature survey. Through the analysis, it is understood that [26] biomedical relationships are done using knowledge-enhanced reading comprehension (KRC) but lead to more misclassification.[27] Clinical relation extraction using knowledge base and BioSentVec embedding is performed but more memory is required for training. [28] uses a machine learning approach for finding a clinical relationship but there are high possibilities of error propagation. [29] Relation extraction is done using a rule-based mining approach but it requires a lot of manual work. [30] Despite requiring extra processing time, an annotation approach for natural language processing is investigated. To get rid of the above-mentioned issues a novel knowledge discovery technique is proposed which is elaborated in the next section.
In order to extract valuable information from clinical text data, clinical entities should be recognized and relations extracted. we have incorporated XLNet, the clinical version of the large bidirectional transformer model, as a basis to resolve discrepancies in previous BERT-based models. This procedure integrates the position embedding into clinical XLNet so that entity markers can be used with the existing [CLS] vector to take advantage of the entities' position data. As a means of taking advantage of the dependent event relation association and the independent events, the multinomial Naïve Bayes probability function has been performed parallel to the joint conditional probability function. A multi-head attention layer is incorporated at the top of the clinical XLNet to refine the relationships between entity pairs. Then, the transformer outputs are concatenated through a linear layer to refine the relations between two entity pairs. Lastly, the softmax classifier can be used to present entity correlation, such as drug-adverse drug event, drug-reason, etc. A graphical or statistical presentation of the absolute and abundant knowledge gleaned from entity relation extraction will be given at the end. The block diagram of the proposed methodology is depicted in Fig. 1.
3.1 Clinical Entity Recognition:
Clinical text data has been adopted in order to extract the knowledge. Pre-processing of the clinical text data was done initially to preserve the prominence of the input corpus to extract knowledge. Tokenization, parts of speech tagging, parsing, stemming, and lemmatization are some of the pre-processing steps.
A critical component of relation extraction is recognizing named entities in clinical settings. It is also important to note that failures in entity recognition were carried forward to the knowledge discovery pipeline's relation extraction task. This leads to error propagation and errors in discovering knowledge, which ultimately results in performance degradation. This issue was avoided by using only one model for clinical entity recognition and relation extraction. The existing approaches rely on sequence labeling models like BERT, Bi-LSTM, and CRF, which suffer from discrepancy issues when tuned with contextual embedding. The XLNet, a large bidirectional transformer model, was chosen as a base model for this research to deal with this issue.
The XLNet has been further fine-tuned by using smaller clinical data so that task-specific knowledge is fully utilized to create a clinical XLNet, where named entities are recognized. Clinical entity recognition is largely aided by the embedding layers. Permutation Language Modeling should be used to pre-train the XLNET embedding with the clinical notes datasets.
The pre-processed data is initially fed into the clinical XLNet model, while the token embedding layer turns the data into tokens for further processing. The [CLS] token has been used as a classification token in clinical XLNet, and the [SEP] token has been used to separate sentences. However, the [CLS] token only uses a portion of the input sequence's knowledge, which has an impact on entity positioning. To address this, position embedding was included in clinical XLNet in order to use the position information of entities by entity markers with the existing [CLS] vector, which helped to carry forward the entire knowledge up to the task's conclusion.
Each entity is saved till the completion of the knowledge discovery process, much like the entity markers. Drug, class, dosage, frequency, duration, route, and condition are all clinical entities. The relationship between the drug and the other entities has thus been retrieved using the relation extraction procedure, which is discussed in the next section.
3.2 Clinical Relation Extraction:
Clinical relation extraction is a crucial task for knowledge discovery from clinical text data. Clinical relation extraction implies the extraction of the relation between the drug with other presented clinical entities like a drug dose, drug frequency, drug-route, drug duration, drug-adversarial drug event, etc. Since we have adopted the n2c2 corpus as clinical text data, there are eight relations that have been identified for extraction, which are listed in Table 2, that also represent the respective description of the clinical entity with the example.
|
Type of Entity |
Description |
Example |
|---|---|---|
|
Drug Name |
Denotes the name of the medicine |
Paracetamol |
|
Drug Class |
Denotes the Chemical class of the medicine |
Analgesics and antipyretics |
|
Dosage |
Dose or concentration of the medicine |
500 mg |
|
Frequency |
Frequency of drug administration |
3 times/day |
|
Duration |
The time range for drug administration |
5 days |
|
Route |
Mode of drug administration |
By mouth |
|
Condition |
The event that provokes the administration |
If aches, pains, and also if fever occurs |
|
Adverse event |
The side effect of the drug |
Rashes |
Even though the XLNet model can overcome the discrepancy issue in the BERT model, it is suffered from extracting dependent clinical events associated with the inbuilt joint conditional probability function alone. In order to solve this, the multinomial Naïve Bayes probability function has been performed in parallel to the joint conditional probability function. Whereas the multinomial Naïve Bayes leverages the prior probability with conditional probability to obtain the likelihood of the dependent event relation association.
The multinomial Naïve Bayes calculates the probabilities of the dependent relations. Let us consider the set of entities as C.Let N be the size of the data, then the MNB allocates the dependent relation
of text m at given entity E. This predicts the relationship between the dependent event association. For example, if the disease entity is fever means, paracetamol must come next as a drug entity, which is called the dependent event association. The joint control probability is used to extract the independent relation from the entities recognized. It uses positional embedding as a way of encoding order within enhanced Clinical data. As a whole, the multinomial Naïve Bayes probability function predicts the dependent clinical event association, as well as the joint conditional probability function, predicts the independent clinical event association.
Figure 2 depicts an example of position information of the clinical entities, in addition to the entity-relationship between drugs and various clinical entities, which are illustrated with different colors.
Furthermore, the relation between the entity pairs, which are presented in the consecutive sentences was unable to be retrieved by the existing models. In the proposed clinical XLNet model, the relation between entity pairs presented in the consecutive sentences is brought under consideration as cross-sentence relation through the incorporation of a multi-head attention layer at the top of the clinical XLNet. The cross sentence relation is identified with the cross entity distance, whereas if the cross entity pair has been identified, the multi-head attention attends on both the sentences to attain the associated relation between the entity pair. Then the clinical XLNet outputs are concatenated through a linear layer. Finally, the softmax classifier can be employed to represent the entity relation, which might be drug-adverse drug event, drug-reason, etc. At last, the absolute and abundant knowledge discovered from the entity relation extraction through text mining with the required relation would be portrayed in a graphical or statistical form. The resulting knowledge has been more beneficial for miners or doctors in decision-making The further implementation particulars of the proposed end-to-end knowledge discovery with the clinical XLNet model have been presented in the next section.
This section provides a detailed explanation of the implementation results as well as the evaluation metrics of the proposed system. A comparison has also been provided to demonstrate that the proposed model's performance is superior to that of existing models.
4.1 Experimental Setup:
This work has been implemented in the working platform of python with the following system specification and the simulation results are discussed below.
4.1.1 Dataset Description:
The corpus utilized in the proposed methodology is National NLP Clinical Challenges (n2c2) 2018 dataset. The n2c2 corpus contains the electronic health record of several patients in 10 classes. The medical history, diagnosis, drug name, class of the drug, treatment plans, vaccination dates, allergies, frequency of the intake of drugs, duration, route, condition, radiological pictures, and laboratory and test results of a patient are all kept in the record. The discharge summaries of patients presented in the corpus are taken into account for knowledge discovery. The n2c2 corpus contains the following entity annotations: drug, strength, form, dosage, frequency, route, duration, reason, ADR. There are eight different kinds of clinical entity relations: strength–drug (severity), form–drug (form), dosage–drug (do), frequency–drug (fr), route–drug (route), duration–- drug (du), reason–drug (reason), ADR–Drug (adverse).
4.2 Simulation Outputs:
This section presents the simulation outputs and their discussion in detail upon knowledge discovery.
The n2c2 corpus has been employed as the input to the proposed clinical XLNet model for knowledge discovery. The corpus is grouped into 10 classes each containing several health records of patients which have been initially pre-processed for cleaning, then represented statistically in Fig. 3.
For knowledge discovery, the clinical entity recognition task is the one, which has to be focused on to enhance the proposed model's efficiency. To do that, the tokenization takes place with token embedding, which is depicted in Fig. 4. On the obtained tokens, the incorporation of position embedding has been identifying the position of the entities as well as the entity relation with the drug, which has been portrayed in Fig. 6. In Fig. 5(a) the entity types marked in the input record have been illustrated. In Fig. 5(b) the identified entity relations with the drug are listed out in the output window with the position of entities, whereas the entity drug is indicated as chemical.
Figure 6 illustrates the output of the proposed model as knowledge discovery. In the figure, the drug-dose relation knowledge has been illustrated from the given input clinical text data for sample representation. Likewise, all eight relations can be extracted by the proposed end-to-end knowledge discovery model. Since we have concentrated on extracting both independent and dependent relation association through multinomial naïve Bayes probability function, our model performed well in relation extraction upon knowledge discovery.
As presented in Fig. 6, the required relation knowledge has been extracted from the input health records, which are graphically presented in Fig. 7 and statistically represented in Table 3, through the mining from the obtained relation extraction. This knowledge was used by the doctors or government for decision-making. For ease of representation, 2000 records have been taken into account in the statistical and graphical representation.
|
No. of Records |
Drug-Dose |
|---|---|
|
200 |
982 |
|
400 |
1064 |
|
600 |
789 |
|
800 |
1325 |
|
1000 |
914 |
|
1200 |
935 |
|
1400 |
796 |
|
1600 |
1023 |
|
1800 |
1244 |
|
2000 |
958 |
4.3 Performance Evaluation Metrics:
In this section, the effectiveness of the proposed method has been assessed using evaluation measures such as accuracy, precision, recall, and F1 score.
Figure 8 represents the overall accuracy of the proposed model, which is also listed in Table 4. The accuracy has been calculated by training the model with a different number of records, whereas the accuracy of the model increased with the number of records and achieves 97.32%. This is because of the restriction of error propagation from named entity recognition into relation extraction.
|
No. of Records |
Accuracy (%) |
|---|---|
|
200 |
97.1 |
|
400 |
97.13 |
|
600 |
97.15 |
|
800 |
97.18 |
|
1000 |
97.2 |
|
1200 |
97.23 |
|
1400 |
97.25 |
|
1600 |
97.28 |
|
1800 |
97.30 |
|
2000 |
97.32 |
Figure 9 represents the overall precision of the proposed model, also recorded in Table 5. From the graph, it is obvious that the suggested system gives high precision of 98.32% with the closeness of accuracy. Like accuracy, the precision is also increased with the number of record showing to the error propagation restriction.
|
No. of Records |
Precision (%) |
|---|---|
|
200 |
98.1 |
|
400 |
98.13 |
|
600 |
98.15 |
|
800 |
98.18 |
|
1000 |
98.2 |
|
1200 |
98.23 |
|
1400 |
98.25 |
|
1600 |
98.28 |
|
1800 |
98.30 |
|
2000 |
98.32 |
The recall of the proposed clinical XLNet model has been depicted in Fig. 10 and values are scheduled in Table 6. Since the proposed model can extract the relation between the entity pair in the consecutive sentences, the performance increases as well up to 98.5% with the increased number of records.
|
No. of Records |
Recall (%) |
|---|---|
|
200 |
98.2 |
|
400 |
98.23 |
|
600 |
98.25 |
|
800 |
98.28 |
|
1000 |
98.4 |
|
1200 |
98.43 |
|
1400 |
98.45 |
|
1600 |
98.48 |
|
1800 |
98.50 |
|
2000 |
98.52 |
Figure 11 and Table 7 clearly show the F1-Score of the proposed model, which is about 98.512 percent. As the number of records increases, the F1-Score of the model also increases. The F1-Score of the proposed system is unwavering by using probabilities utilized in the proposed approach.
|
No. of Records |
F1-Score (%) |
|---|---|
|
200 |
98.21 |
|
400 |
98.213 |
|
600 |
98.215 |
|
800 |
98.218 |
|
1000 |
98.41 |
|
1200 |
98.413 |
|
1400 |
98.415 |
|
1600 |
98.418 |
|
1800 |
98.510 |
|
2000 |
98.512 |
4.4 Comparison Analysis:
In this section, various performances of the proposed clinical XLNet model have been compared with the existing models for clinical knowledge discovery through relation extraction to ensure the performance of the proposed model. The evaluation metrics such as accuracy, precision, recall as well as F1-score were compared with the existing models such as BERT [30],Syntax convolutional neural network (SCNN) [31], CNN-bioWE [32], Multichannel convolutional neural network (MCCNN) [33], Joint AB-LSTM [34], Recursive neural network (RvNN) [35], Position-aware LSTM[36] models.
The accuracy of clinical knowledge discovery from clinical text data through the proposed clinical XLNet has been compared with the existing clinical models, which is demonstrated in Fig. 12 and statistically presented in Table 8. From the graph, it is understood that the transformer models would be performed well in the clinical task than other CNN and LSTM based techniques, however, our proposed XLNet is performed well than BERT based model and achieved 21.2% higher accuracy than the BERT model.
|
Techniques |
Accuracy (%) |
|---|---|
|
SCNN |
70 |
|
CNN-bioWE |
74 |
|
MCCNN |
74.1 |
|
Joint AB-LSTM |
72 |
|
RVNN |
74.6 |
|
Position-aware LSTM |
75 |
|
BERT |
76.1 |
|
Proposed |
97.32 |
The precision of clinical knowledge discovery from clinical text data via the proposed clinical XLNet was compared to existing clinical models, as shown in Fig. 13 and quantitatively compiled in Table 9. The transformer models appear to perform better in clinical tasks, yet the proposed XLNet outperforms the BERT model by 22 percent.
|
Techniques |
Precision (%) |
|---|---|
|
SCNN |
69.1 |
|
CNN-bioWE |
75.7 |
|
MCCNN |
75.9 |
|
Joint AB-LSTM |
73.4 |
|
RVNN |
74.4 |
|
Position-aware LSTM |
75.8 |
|
BERT |
76.8 |
|
Proposed |
98.86 |
The recall of clinical knowledge discovery from clinical text data through the proposed clinical XLNet has been compared with the existing clinical models, which is exemplified in Fig. 14 and statistically presented in Table 10. From the graph, it is understood that the transformer models would be performed well in the clinical task than other CNN and LSTM based techniques, however, our proposed XLNet is performed well than BERT based model and achieved 26.5% higher than the BERT model.
|
Techniques |
Recall (%) |
|---|---|
|
SCNN |
65.1 |
|
CNN-bioWE |
64.7 |
|
MCCNN |
65.2 |
|
Joint AB-LSTM |
69.6 |
|
RVNN |
69.3 |
|
Position-aware LSTM |
70.4 |
|
BERT |
71.3 |
|
Proposed |
97.86 |
The F1-Score of the proposed clinical XLNet of clinical knowledge acquisition from clinical text data has been compared to existing clinical models, as shown in Fig. 15 and quantitatively given in Table 11. The transformer models will perform better in the clinical job than other CNN and LSTM-based strategies, as depicted in the graph; nevertheless, our proposed XLNet outperforms the BERT-based model by 13.7 percent in terms of F1-Score.
|
Techniques |
F1-Score (%) |
|---|---|
|
SCNN |
67 |
|
CNN-bioWE |
69.8 |
|
MCCNN |
70.2 |
|
Joint AB-LSTM |
71.5 |
|
RVNN |
71.7 |
|
Position-aware LSTM |
73 |
|
BERT |
73.9 |
|
Proposed |
97.3 |
In this paper, a novel end-to-end clinical knowledge discovery strategy has been proposed with clinical XLNet, whereas, the clinical entity recognition, as well as relation extraction, were performed solely on the proposed model to restrict the error propagation. In order to leverage the dependent event relation association along with the independent events, the multinomial Naïve Bayes probability function has been performed in parallel to the joint conditional probability function. Additionally, the proposed model has taken into account the extraction of entity pairs in consecutive sentences. This provides vast knowledge from the clinical text data, which has been implemented with the n2c2 corpus. The results reveal that the proposed model performs well with an accuracy of 97.32%, precision of98.86%, recall of 97.86%, and F1-score of97.3%.
- Human and Animal Ethics
Not Applicable
- Authors' contributions
All the authors contributed to the design and implementation of the research, to the analysis of the results, and to the writing of the manuscript.
all authors' individual contributions,
Conceptualization; 1 Naveen S Pagad
Data curation; 2 Pradeep N
Formal analysis; 3 Tulika Chakrabarti, 9Samuel Frimpong
Investigation: ; 4 Prasun Chakrabarti
Methodology; 5*Ahmed A. Elngar
Project administration; 6Martin Margala
Resources; 7Mohammad Nami
Software; 8Neha Sharma
Supervision; 1 Naveen S Pagad
Validation; 2 Pradeep N
Visualization; 3 Tulika Chakrabarti
Roles/Writing - original draft: 4 Prasun Chakrabarti
Writing - review & editing, 5*Ahmed A. Elngar
- Acknowledgments
The corresponding author is thankful to the co-authors for their support to complete the manuscript. We are also thankful to the anonymous reviewers and Editors for their perceptive comments on the manuscript.
Ethical Approval
The manuscript is original and has not been published, accepted for publication or under editorial review for publication elsewhere.
Consent to Participate
Not applicable
Consent to Publish
Not applicable
Funding
No funding was received to assist with the preparation of this manuscript.
Availability of data and materials
All data generated or analyzed as part of this study are included in this published article.
Competing Interests
☒ The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
☒The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:
There is No known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
- Keshta, Ismail; Odeh, Ammar (2020). Security and privacy of electronic health records: Concerns and challenges. Egyptian Informatics Journal, (), S1110866520301365–.
- Moghaddasi, Hamid, Azam Sadat Hosseini, and FarkhondehAsadi. "Features and Services of Well-designed Hospital Information Systems: A Review Study." Archives of Advances in Biosciences 12.2 (2021): 55-66.
- Kiwanuka, Achilles, Moses Bagyendera, Joseph Wamema, Andrew AlunyuEgwar, Mercy Amiyo, Andrew Kambugu, and Josephine Nabukenya. "Establishing the State of Practice about Data Standards in Monitoring Healthcare Interventions for HIV in Uganda's EMR-based Health Information Systems." In HEALTHINF, pp. 200-211. 2021.
- Miller, M. J., Xu, L., Qin, J., Hahn, E. E., Ngo-Metzger, Q., Mittman, B., ... & Chao, C. R. (2021). Impact of COVID-19 on cervical cancer screening rates among women aged 21–65 years in a large integrated health care system—Southern California, January 1–September 30, 2019, and January 1–September 30, 2020. Morbidity and Mortality Weekly Report, 70(4), 109.
- Rashid, R., Chang, C., Niu, F., Deguzman, L., Draves, M., Awsare, S., & Hui, R. (2020). Evaluation of a pharmacist-managed nonsteroidal anti-inflammatory drugs deprescribing program in an integrated health care system. Journal of Managed Care & Specialty Pharmacy, 26(7), 918-924.
- Juhn, Young; Liu, Hongfang (2019). Natural language processing to advance EHR-based clinical research in Allergy, Asthma, and Immunology. Journal of Allergy and Clinical Immunology, (), S0091674919326041–.
- Olivetti, Elsa A.; Cole, Jacqueline M.; Kim, Edward; Kononova, Olga; Ceder, Gerbrand; Han, Thomas Yong-Jin; Hiszpanski, Anna M. (2020). Data-driven materials research enabled by natural language processing and information extraction. Applied Physics Reviews, 7(4), 041317–
- Topalovic, Marko, et al. "Artificial intelligence outperforms pulmonologists in the interpretation of pulmonary function tests." European Respiratory Journal 53.4 (2019).
- Attia, Zachi I., et al. "Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram." Nature medicine 25.1 (2019): 70-74.
- Mercan, Ezgi, et al. "Assessment of machine learning of breast pathology structures for automated differentiation of breast cancer and high-risk proliferative lesions." JAMA network open 2.8 (2019): e198777-e198777.
- Kong, J., Zhang, L., Jiang, M., & Liu, T. (2021). Incorporating multi-level CNN and attention mechanism for Chinese clinical named entity recognition. Journal of Biomedical Informatics, 116, 103737.
- Ravikumar, J., and Kumar P. Ramakanth. "Machine learning model for clinical named entity recognition." International Journal of Electrical and Computer Engineering 11.2 (2021): 1689.
- Xie, B., Shen, G., Guo, C., & Cui, Y. (2021). The Named Entity Recognition of Chinese Cybersecurity Using an Active Learning Strategy. Wireless Communications and Mobile Computing, 2021.
- Lee, Lung-Hao, and Yi Lu. "Multiple Embeddings Enhanced Multi-Graph Neural Networks for Chinese Healthcare Named Entity Recognition." IEEE Journal of Biomedical and Health Informatics (2021).
- Yaseen U, Langer S. Neural Text Classification and StackedHeterogeneousEmbeddings for Named Entity Recognition in SMM4H 2021. arXiv preprint arXiv:2106.05823. 2021 Jun 10.
- Hung, J. C., & Chang, J.-W. (2021). Multi-level transfer learning for improving the performance of deep neural networks: Theory and practice from the tasks of facial emotion recognition and named entity recognition. Applied Soft Computing, 109, 107491.
- Fu, Yingwen, et al. "An Open-Source Dataset and A Multi-Task Model for Malay Named Entity Recognition." arXiv preprint arXiv:2109.01293 (2021).
- Sarzynska-Wawer J, Wawer A, Pawlak A, Szymanowska J, Stefaniak I, Jarkiewicz M, Okruszek L. Detecting formal thought disorder by deep contextualized word representations. Psychiatry Research. 2021 Oct 1;304:114135.
- Jia L, Liu S, Wei F, Kong B, Wang G. Nested Named Entity Recognition via an Independent-Layered Pretrained Model. IEEE Access. 2021 Aug 5;9:109693-703.
- Xu, Y., Huang, H., Feng, C., & Hu, Y. (2021, May). A Supervised Multi-Head Self-Attention Network for Nested Named Entity Recognition. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 16, pp. 14185-14193).
- Straková, Jana, Milan Straka, and Jan Hajič. "Neural architectures for nested NER through linearization." arXiv preprint arXiv:1908.06926 (2019).
- Mahendran, Darshini, and Bridget T. McInnes. "Extracting Adverse Drug Events from Clinical Notes." arXiv preprint arXiv:2104.10791 (2021).
- Chapman, Alec B., et al. "Detecting adverse drug events with rapidly trained classification models." Drug safety 42.1 (2019): 147-156.
- Dandala, Bharath, VenkataJoopudi, and Murthy Devarakonda. "Adverse drug events detection in clinical notes by jointly modeling entities and relations using neural networks." Drug safety 42.1 (2019): 135-146.
- Gangwar, A., Jain, S., Sourav, S., & Modi, A. (2021). Counts@ IITK at SemEval-2021 Task 8: SciBERT Based Entity And Semantic Relation Extraction For Scientific Data. arXiv preprint arXiv:2104.01364.
- Chen, J., Hu, B., Peng, W. et al. Biomedical relation extraction via knowledge-enhanced reading comprehension. BMC Bioinformatics 23, 20 (2022). https://doi.org/10.1186/s12859-021-04534-5
- Alimova, Ilseyar, and Elena Tutubalina. "Multiple features for clinical relation extraction: a machine learning approach." Journal of biomedical informatics 103 (2020): 103382.
- Hong, Lixiang, et al. "A novel machine learning framework for automated biomedical relation extraction from large-scale literature repositories." Nature Machine Intelligence 2.6 (2020): 347-355.
- Alodadi, Mohammad S., and Vandana P. Janeja. "Linking Knowledge Discovery In Clinical Notes And Massive Biomedical Literature Repositories." 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2019.
- Piad-Morffis, Alejandro, et al. "A general-purpose annotation model for knowledge discovery: Case study in spanish clinical text." Proceedings of the 2nd Clinical Natural Language Processing Workshop. 2019.
- Zhao, Z., Yang, Z., Luo, L., Lin, H. & Wang, J. “Drug drug interaction extraction from biomedical literature using syntax convolutional neural network”. Bioinformatics 32, 3444–3453 (2016)
- Liu, S., Tang, B., Chen, Q. & Wang, X. Drug-drug interaction extraction via convolutional neural networks. Comput. Math. Methods Med. 2016, 6918381 (2016).
- Quan, C., Hua, L., Sun, X. & Bai, W. Multichannel convolutional neural network for biological relation extraction. Biomed Res. Int. 2016, 1850404 (2016).
- Sahu, S. K. &Anand, A. Drug–drug interaction extraction from biomedical texts using long short-term memory network. J. Biomed. Inform. 86, 15–24 (2018).
- Lim, S., Lee, K. & Kang, J. Drug drug interaction extraction from the literature using a recursive neural network. PLoS ONE 13, e0190926 (2018).
- Zhou, D., Miao, L. & He, Y. Position-aware deep multi-task learning for drug–drug interaction extraction. Artif. Intell. Med. 87, 1–8 (2018).
No competing interests reported.