Clinical report classification using Natural Language Processing and Topic Modeling.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Electronic health records contain vast amounts of unstructured clinical text.
• Efficiently processing and coding this text is crucial for medical decision-making.
• Accurate classification of Emergency Department (ED) computed tomography (CT) reports is needed.

Purpose of the Study:

• To investigate techniques for improving the classification of ED CT reports.
• To develop a system using NLP and machine learning for coding clinically important injuries in traumatic orbital fracture victims.
• To explore topic modeling as an alternative text representation for classification.

Main Methods:

• Utilized Natural Language Processing (NLP) to extract structured information from free-text CT reports.
• Applied machine learning techniques for classifying the presence of injuries.
• Employed topic modeling to represent patient reports in a lower-dimensional space.
• Investigated NLP feature filtering using modifiers and binary topic classification.

Main Results:

• Both NLP and topic modeling approaches improved raw text classification performance.
• Filtering NLP codes with modifiers yielded the best classification results.
• Topic modeling provided effective dimensionality reduction with comparable classification accuracy to NLP features.
• Binary topic classification did not outperform raw text classification.

Conclusions:

• NLP techniques, particularly with modifier filtering, are effective for enhancing CT report classification.
• Topic modeling offers a viable alternative for dimensionality reduction and classification of clinical text.
• Further research may optimize topic modeling for improved clinical data analysis.