Advanced AI Techniques for ICD Code Prediction

In my latest research, I co-authored a paper titled “Co-Occurrence Graph-Enhanced Hierarchical Prediction of ICD Codes,” which was presented at ICASSP 2024. This work demonstrates my expertise in designing and developing complex AI/NLP models, particularly in the context of multi-label classification problems.

Summary of the Research:
Objective:
The goal of this research was to enhance the accuracy of automated ICD coding by integrating graph-based relationships of ICD code co-occurrence with hierarchical text representations derived from the ICD ontology.

Key Contributions:
Modular Approach: The study introduced a modular approach combining graph-based integration of ICD code co-occurrence and hierarchical-enriched text representations. This approach significantly improved the accuracy of ICD code predictions.
Graph-Based Integration: A novel one-layer graph-based model was developed to explore high-order interactions among ICD codes within the dataset, enriching the text representation.
Hierarchical Enhancement: The hierarchical structure of ICD codes was leveraged to enhance text features, improving model performance.
Methodology:
Text Representation Learning: Using a Convolutional Neural Network (CNN) for initial text representation, the CAML architecture was employed for its simplicity and computational efficiency.
Hierarchical Feature Enhancement: Attention mechanisms were applied at different hierarchical levels of the ICD ontology to refine features progressively.
Graph-Based Module: A graph convolutional layer was integrated to capture the co-occurrence relationships between ICD codes, further enriching the feature representation.
[Placeholder for a workflow or pipeline image]

Results:
Our experiments demonstrated substantial performance gains using the combined approach. The model with both hierarchical and graph-based enhancements outperformed baseline models, highlighting the importance of integrating external knowledge sources in improving ICD coding accuracy.

Performance Metrics: Significant improvements were observed in key metrics such as f1-micro, f1-macro, and precision@8, showcasing the effectiveness of our approach.
Conclusion:
This research underscores my ability to design and implement sophisticated AI/NLP models that leverage both hierarchical and graph-based data structures to enhance predictive accuracy. The modular design of the proposed system allows for seamless integration into existing AI frameworks, demonstrating its practical applicability across various domains.