Semi-self-supervised Automated ICD Coding

  title={Semi-self-supervised Automated ICD Coding},
  author={Hlynur Dav'idh Hlynsson and Steind{\'o}r Ellertsson and J{\'o}n Friðrik Daðason and Emil Larus Sigurdsson and Hrafn Loftsson},
Clinical Text Notes (CTNs) contain physicians’ 001 reasoning process, written in an unstructured 002 free text format, as they examine and inter003 view patients. In recent years, several studies 004 have been published that provide evidence for 005 the utility of machine learning for predicting 006 doctors’ diagnoses from CTNs, a task known 007 as ICD coding. Data annotation is time con008 suming, particularly when a degree of special009 ization is needed, as is the case for medical 010 data… 

Figures and Tables from this paper


Towards BERT-based Automatic ICD Coding: Limitations and Opportunities
This work investigates in detail ICD coding using PubMedBERT, a state-of-the-art transformer model for biomedical language understanding, and finds that the difficulty of fine-tuning the model on long pieces of text is the main limitation for BERT-based models on I CD coding.
BERT-XML: Large Scale Automated ICD Coding Using BERT Pretraining
A machine learning model, BERT-XML, is proposed for large scale automated ICD coding of EHR notes, utilizing recently developed unsupervised pretraining that have achieved state of the art performance on a va-riety of NLP tasks.
Evaluation and accurate diagnoses of pediatric diseases using artificial intelligence
This study shows that MLCs can query EHRs in a manner similar to the hypothetico-deductive reasoning used by physicians and unearth associations that previous statistical methods have not found, and provides a proof of concept for implementing an AI-based system to aid physicians in tackling large amounts of data, augmenting diagnostic evaluations, and to provide clinical decision support in cases of diagnostic uncertainty or complexity.
On the Contribution of Per-ICD Attention Mechanisms to Classify Health Records in Languages with Fewer Resources than English
A multi-label text classifier with per-label attention for the classification of Electronic Health Records according to the International Classification of Diseases, which leverages the BERT Multilingual model to share the language modelling capabilities across the languages.
Artificial intelligence in the GPs office: a retrospective study on diagnostic accuracy
Light is shed on the diagnostic accuracy of ML in a clinical context, as well as the interpretation of its predictions, if the vast potential of ML is to be utilized in primary health care, its performance, safety, and inner workings need to be understood by clinicians.
A survey on semi-supervised learning
This survey aims to provide researchers and practitioners new to the field as well as more advanced readers with a solid understanding of the main approaches and algorithms developed over the past two decades, with an emphasis on the most prominent and currently relevant work.
A Systematic Literature Review of Automated ICD Coding and Classification Systems using Discharge Summaries
Computerised systems that employ Artificial Intelligence, Machine Learning, Deep Learning and Natural Language Processing for assigning ICD codes to discharge summaries are reviewed.
A Warm Start and a Clean Crawled Corpus - A Recipe for Good Language Models
It is demonstrated that a properly cleaned crawled corpus is sufficient to achieve state-of-the-art results in NLP applications for low to medium resource languages, by comparison with models trained on a curated corpus.
The advantages of the Matthews correlation coefficient (MCC) over F1 score and accuracy in binary classification evaluation
This article shows how MCC produces a more informative and truthful score in evaluating binary classifications than accuracy and F1 score, by first explaining the mathematical properties, and then the asset of MCC in six synthetic use cases and in a real genomics scenario.
ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators
The contextual representations learned by the proposed replaced token detection pre-training task substantially outperform the ones learned by methods such as BERT and XLNet given the same model size, data, and compute.