Natural Language Processing for Clinical Laboratory Data Repository Systems: Implementation and Evaluation for...

Area of Science:

• Medical Informatics
• Computational Biology
• Artificial Intelligence

Background:

• Laboratory data repositories are growing in volume and complexity, making manual data extraction for secondary uses challenging.
• Natural Language Processing (NLP) offers automated solutions for extracting clinically meaningful information from unstructured text.

Purpose of the Study:

• To evaluate a deep learning-based NLP model as an alternative to resource-intensive rule-based systems for information extraction.
• To develop and assess an NLP model for extracting knowledge from text-based laboratory reports.

Main Methods:

• A hierarchical multilabel classifier NLP model was trained on 87,500 laboratory reports for 14 respiratory viruses.
• The model was trained to classify reports into 24 fine-grained and 6 coarse-grained labels.
• Performance was evaluated for stability, variation, and generalizability across internal and external test sets.

Main Results:

• The NLP model achieved microaveraged F1-scores >94% across all classes on internal, historical, and external test sets.
• Higher precision and recall were observed on internal and historical data, with performance variations due to data imbalance.
• The model's performance was lower for virus detection cases (lowest F1-score of 57%) compared to testing cases.

Conclusions:

• Deep learning NLP models show significant promise for extracting information from laboratory reports.
• These models provide scalable, timely, and practical access to high-quality encoded laboratory data.
• Integration into laboratory information systems can enhance data accessibility and utility.