Applying Convolutional Neural Networks to Predict the ICD-9 Codes of Medical Records

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• The International Statistical Classification of Disease and Related Health Problems (ICD) is crucial for disease reporting.
• Existing disease prediction systems often require extensive clinical data from healthcare professionals.
• Patient-reported information, particularly subjective notes, is underutilized in automated diagnostic coding.

Purpose of the Study:

• To develop and evaluate a deep learning algorithm for disease classification and diagnostic code prediction.
• To utilize only the subjective component of medical progress notes for prediction.
• To assess the feasibility of a self-diagnosis assistance tool based on patient self-reports.

Main Methods:

• A dataset of approximately 168,000 medical records (2003-2017) was used.
• Standard text processing and word embedding techniques were applied to patient progress notes.
• A convolutional neural network (CNN) model was developed to predict ICD-9 codes from subjective notes.

Main Results:

• The model achieved an overall accuracy of 0.409, recall of 0.409, and precision of 0.436 for ICD-9 code prediction.
• When considering only "chapter match" for ICD-9 codes, performance improved to an accuracy of 0.580, recall of 0.580, and precision of 0.582.
• The model demonstrated the effectiveness of using subjective patient descriptions for diagnostic coding.

Conclusions:

• Deep learning models can effectively predict diagnostic codes using only subjective patient self-reports.
• This approach offers a potential low-resource solution for preliminary remote diagnosis and triage.
• The findings support the development of accessible health assistance tools, particularly for underserved areas.