OLR-Net: Object Label Retrieval Network for principal diagnosis extraction.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automated extraction of principal diagnoses from patient discharge summaries is crucial for medical data utilization.
• Manual extraction is time-consuming; existing automatic models struggle with rare diagnoses due to limited training data.

Purpose of the Study:

• To develop a method for extracting principal diagnoses from discharge summaries, specifically addressing scenarios with limited available data.

Main Methods:

• Proposed the Object Label Retrieval Network (OLR-Net) incorporating semantic extraction, label localization, and label retrieval.
• Utilized one-dimensional convolutional neural networks for diagnosis localization to improve performance on rare diagnoses.
• Evaluated performance using hit ratio, mean reciprocal rank, and area under the receiver operating characteristic curve (AUROC).

Main Results:

• Tested on 12,788 oncology discharge summaries across five data settings (full, top-50, few-shot, one-shot, zero-shot).
• Achieved a high HR@5 of 0.8778 and macro-AUROC of 0.9851 with the full dataset.
• Demonstrated strong performance in limited data settings, with macro-AUROC of 0.9833 for few-shot and 0.9485 for one-shot data.

Conclusions:

• The OLR-Net model shows significant potential for principal diagnosis extraction from discharge summaries, particularly when data is limited.
• Label localization and retrieval are key components enabling effective extraction in low-data scenarios.