A Scalable Radiomics- and Natural Language Processing-Based Machine Learning Pipeline to Distinguish Between Painful...

Area of Science:

• Oncology
• Radiology
• Medical Informatics
• Artificial Intelligence

Background:

• Objective pain biomarkers are crucial for understanding cancer pain prognosis and management.
• Artificial intelligence (AI) can identify pain biomarkers in patients with bone metastases (BMs).

Purpose of the Study:

• Develop and evaluate a scalable AI pipeline using natural language processing (NLP) and radiomics.
• Differentiate between painful and painless BM lesions in CT images using imaging features.

Main Methods:

• Retrospective study of 176 patients with thoracic spine BMs.
• NLP extracted physician-reported pain scores from clinical notes.
• Radiomics features extracted from CT-based regions of interest (ROIs).
• Machine learning models evaluated using standard performance metrics.

Main Results:

• A neural network classifier achieved an area under the receiver operating characteristic curve of 0.83 in the test set.
• The best model demonstrated 82% accuracy and 85% specificity.
• The pipeline successfully differentiated painful from painless BM lesions.

Conclusions:

• The developed NLP and radiomics pipeline effectively distinguishes painful from painless BM lesions.
• The method is scalable, leveraging NLP for pain scores and CT imaging for BM identification.