Machine Learning Algorithms to Predict Breast Cancer Recurrence Using Structured and Unstructured Sources from...

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Breast cancer (BC) recurrence significantly impacts patient mortality.
• Machine learning (ML) offers potential for improved patient risk stratification using healthcare data.
• Combining structured and unstructured data sources is explored for enhanced BC recurrence prediction.

Purpose of the Study:

• To evaluate the effectiveness of ML algorithms in predicting 5-year breast cancer recurrence.
• To compare prediction performance using structured data, unstructured (free text) data, and a combination of both.
• To identify the optimal data source and ML model for breast cancer recurrence risk stratification.

Main Methods:

• Collected and preprocessed clinical data from 823 breast cancer patients.
• Derived features from structured clinical information and unstructured free-text clinical notes.
• Evaluated five ML algorithms, including XGBoost, for predicting 5-year recurrence.
• Assessed model performance using precision, recall, F1-score, and AUROC.

Main Results:

• The XGBoost model demonstrated the highest performance (AUROC = 0.807).
• Structured data yielded the best prediction results, followed by unstructured data.
• The combined dataset (structured and unstructured) performed the poorest.

Conclusions:

• ML algorithms are valuable for breast cancer recurrence risk stratification and patient monitoring.
• Structured clinical data provides superior performance for ML-based recurrence prediction.
• Natural language processing approaches offer comparable results to structured data with potentially less preprocessing effort.