Automated ICD-9 Coding via A Deep Learning Approach.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Linguistics

Background:

• Manual ICD-9 coding is resource-intensive and prone to errors.
• Accurate automated coding is crucial for healthcare efficiency and data analysis.
• Deep learning shows promise in complex classification tasks.

Purpose of the Study:

• To develop and evaluate DeepLabeler, a novel deep learning framework for automated ICD-9 code assignment.
• To compare DeepLabeler's performance against traditional machine learning methods.
• To analyze the contribution of different components within the DeepLabeler framework.

Main Methods:

• Implementation of DeepLabeler, combining Convolutional Neural Networks (CNNs) with 'Document to Vector' (Doc2Vec) for feature extraction.
• Training and evaluation on MIMIC-II and MIMIC-III datasets.
• Comparative analysis with hierarchy-based Support Vector Machines (SVM) and flat SVM models.

Main Results:

• DeepLabeler achieved state-of-the-art performance, with micro F-measures of 0.335 on MIMIC-II and 0.408 on MIMIC-III.
• Outperformed classical SVM methods by at least 14% on both datasets.
• Analysis confirmed CNNs as the most effective component, with Doc2Vec essential for global feature extraction.

Conclusions:

• Deep learning, particularly the DeepLabeler framework, offers a highly effective solution for automated medical coding.
• The combination of CNNs and Doc2Vec provides robust feature representation for text multi-label classification.
• DeepLabeler demonstrates significant potential to enhance the efficiency and accuracy of ICD-9 coding in clinical practice.