Towards automated clinical coding | JoVE Visualize

Area of Science:

Medical Informatics
Computational Linguistics
Machine Learning

Background:

Clinical coding translates patient encounters into standardized codes, essential for healthcare management and research.
Manual coding is resource-intensive, error-prone, and slow, highlighting the need for automated solutions.
Automated clinical coding faces challenges due to complex clinical text and the vast, imbalanced nature of disease codes.

Purpose of the Study:

To explore methods for representing clinical text and hierarchical coding ontologies for improved automated clinical coding.
To evaluate the impact of different text and label representation strategies on coding accuracy.
To leverage hierarchical medical knowledge within statistical models for enhanced disease prediction.

Main Methods:

Clinical text was represented using term frequency-inverse document frequency (TF-IDF) and word embeddings for recurrent neural networks (RNNs).
Disease labels were represented atomically and by composing representations from nodes within a hierarchical coding ontology.
The Medical Information Mart for Intensive Care III (MIMIC-III) dataset was used to predict International Classification of Diseases, ninth revision, Clinical Modification (ICD-9-CM) codes.

Main Results:

Composing label representations from ontology nodes improved weighted F1 scores for predicting 17,561 disease labels (0.264-0.281) compared to atomic representations (0.232-0.249).
RNN text representation improved weighted F1 for 19 disease-category labels (0.682-0.701) over TF-IDF (0.662-0.682).
TF-IDF outperformed RNNs for predicting the 17,561 individual disease labels.

Conclusions:

Incorporating hierarchically structured medical knowledge into statistical models significantly improves automated clinical coding performance.
Performance gains are primarily attributed to better representation of rarer diseases.
While RNNs enhance medical text representation in some contexts, representing very rare diseases remains a challenge for future work.