Towards automatic encoding of medical procedures using convolutional neural networks and autoencoders

Area of Science:

• Medical Informatics
• Computational Linguistics
• Machine Learning

Background:

• Clinical classification systems like ICD-10 and CHOP are vital for healthcare management and data exchange.
• Traditional manual or rule-based classification methods face limitations due to restricted vocabulary and handcrafted rules.
• Conventional machine learning requires laborious, error-prone human-annotated datasets.

Purpose of the Study:

• To present a novel neural network pipeline for automated medical procedure classification.
• To evaluate the effectiveness of autoencoder and convolutional neural network (CNN) models for query-category relevance determination in medical coding.
• To compare these deep learning approaches against traditional machine learning baselines.

Main Methods:

• Developed a processing pipeline utilizing convolutional neural networks (CNNs) and autoencoders with logistic regression.
• Applied the pipeline to the task of medical procedure classification, focusing on relevance determination between query and category text.
• Compared performance against Support Vector Machine (SVM) and logistic regression baselines using various configurations.

Main Results:

• The autoencoder-based method achieved a micro F1 score of 70.29% for relevance determination.
• The CNN-based method reached a micro F1 score of 60.86%, demonstrating high efficiency.
• Both neural network approaches showed promise for automatic encoding compared to baselines.

Conclusions:

• Neural network-based methods, particularly autoencoders, offer a more efficient and potentially more accurate approach to automatic medical procedure classification.
• The developed pipeline overcomes limitations of traditional methods by leveraging deep learning for improved clinical coding.
• Further investigation into advantages and limitations of these models is warranted for practical implementation.