Page 29 - AIH-1-1

Artificial Intelligence in Health NLP in EHR

Table 2. (Continued)
Article Focus NLP level Limitations Future scope
[55] Surveillance of cervical Lexical and semantic Reports are taken from a single health-care Research to study the portability of
cancer: anal cancer and system, and classification is done at the data among institutes to address the
pre-cancer document level, not at the patient level problem of misclassification in EHR
[56] Clinical named entity Phonetic and semantic Only using character-based NER Use of word-based NER
recognition
[58] Lesion summarization and Discourse A small dataset from a single institution Increase training test set and
cancer response generalizability; adoption of an
attention-based method
[59] Identification of Kawasaki Pragmatic Limited variety of syntax, spelling errors, The incorporation of data from its
disease hypothetical clauses, and the syntax for which timestamp analysis and type of EHR
the tool has not trained affects the feasibility of the KD-NLP
tool
[60] Autodetection using EMR Lexical No information extraction from images Improvement in prediction
[61] Extraction of numerical Lexical Abbreviations are reviewed manually, overfitting Requirement of additional keywords
data due to additions, the decision on the variable to extend the model for hospitals
boundary, and the difference between the format
of clinical notes in hospital
[62] Extract ion of the index of Lexical and semantic Selection of reports, a small number of negative Inclusion of more negative cases,
lesions cases, no direct extraction of lesion images direct extraction of lesion images
[63] Identification of cancer Semantic Evaluation to a gold standard Joint model to share weights of both
relate concept identification and relation
extraction
[64] Detection of bleeding events Semantic The gold dataset was relatively small, Topics for further study are bleeding
generalizability and robustness are limited causes, anatomic sites of bleeding,
to a small size, and the system is based on bleeding severity, and assertions from
sentence-level classification but not context EHR.
[29] Identification of patients Lexical User familiarity with implementation Application to an extensive database,
with rheumatoid arthritis software, language-specific preprocessing, and test with a computational language
performance can be further optimized expert, and need of encryption of
clinical notes
[65] Identification of peripheral Semantic Data retrieved from single institutions Validation of tools in other domains,
arterial disease automatic identification of PAD at
point of care
[66] Extraction of left ventricular Lexical, False positive and false harmful errors NA
ejection fraction morphological and
syntactic
[67] Capturing various Lexical and semantic More significant agreement is required for Deriving symptoms domain
dimensions of assessment and cohort generation, portability
neuropathology of data
[68] Extraction of information Lexical and semantic Small annotated set, generalization of conclusion, Generalized solution for data from
from clinical notes use of singular/plural forms only various domains, considering
events tags for annotation process,
rule-based approach
[69] Identification of incidental Syntactic Requirement of manual review in small amounts, Use of machine learning approach to
pulmonary nodules non-specific ambiguous terminologies improve performance
[70] Extracting information Lexical and semantic Inconsistency in annotation; a small set of Generalization of the model, study of
about geriatric syndrome patients; use of limited geriatric constructs temporal patterns of construct
[71] Detection of symptoms of Morphological and Weakness of algorithm due to misclassification Qualitative description of severity,
infectious diseases semantic implementation of large-size
data, use of an algorithm in other
domains, the flexibility of dictionary,
prediction of future symptoms
(Cont’d...)

Volume 1 Issue 1 (2024) 23 https://doi.org/10.36922/aih.2147