Artificial Intelligence Learning Semantics via External Resources for Classifying Diagnosis Codes in Discharge Notes

Area of Science:

• Medical Informatics
• Computational Linguistics
• Machine Learning

Background:

• Automated disease code classification from free-text medical data is crucial for public health surveillance.
• Traditional natural language processing (NLP) pipelines have limitations in accurately classifying disease codes.

Purpose of the Study:

• To compare the performance of traditional NLP pipelines against a novel method combining word embedding and a convolutional neural network (CNN).
• To evaluate the effectiveness of these methods in classifying International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) diagnosis codes from discharge notes.

Main Methods:

• Two classification methods were employed: traditional NLP with supervised machine learning models and a CNN utilizing word embeddings.
• The methods were used to identify chapter-level ICD-10-CM diagnosis codes in 103,390 discharge notes.
• Evaluation used receiver operating characteristic curves, calculating the area under the curve (AUC) and F-measure.

Main Results:

• The proposed word embedding and CNN method achieved higher accuracy (mean AUC 0.9696, mean F-measure 0.9086) than traditional NLP approaches (mean AUC range 0.8183-0.9571).
• A real-world simulation confirmed the superior performance of the proposed method (mean AUC 0.9645, mean F-measure 0.9003).
• CNN's convolutional layers effectively identified keywords and extracted concepts for accurate diagnosis code prediction.

Conclusions:

• The combination of word embedding and CNN demonstrates superior performance and requires minimal data preprocessing compared to traditional methods.
• This advancement reduces limitations posed by incomplete dictionaries and facilitates the extraction of unstructured medical information.
• The findings suggest a future shift towards big data analytics in healthcare through automated information extraction.