Robust Estimation of Breast Cancer Incidence Risk in Presence of Incomplete or Inaccurate Information

Area of Science:

• Medical Informatics
• Machine Learning
• Biostatistics

Background:

• Accurate breast cancer risk estimation is crucial for early detection and prevention.
• Incomplete or inaccurate patient data presents a significant challenge for predictive models.
• Evaluating the robustness of machine learning classifiers is essential for clinical applications.

Purpose of the Study:

• To assess the performance of logistic regression (LR), Random Forests (RF), and a custom Neural Network (NN) in estimating breast cancer risk.
• To determine the robustness of these classifiers when faced with missing or corrupted data.

Main Methods:

• Utilized open data from the Breast Cancer Surveillance Consortium (BCSC) Data Resource.
• Conducted ablation studies by systematically introducing missing (pm) and corrupted (pc) data.
• Compared the discriminative performance (Area Under Curve - AUC) of LR, RF, and NN classifiers under varying data quality conditions.

Main Results:

• The custom NN achieved the highest AUC (0.649) on complete data, followed closely by LR (0.645) and RF (0.643).
• The NN demonstrated superior performance over LR and RF when less than 50% of the data was missing or inaccurate (pm < 0.5, pc < 0.5).
• For data missing/inaccurate above 50%, LR performance became comparable to the custom NN, while RF showed consistently poorer performance.

Conclusions:

• The custom NN provides reliable breast cancer risk estimates, even with significant amounts of missing or inaccurate input data.
• This finding is particularly relevant for healthcare applications where complete individual participant data may not always be available.
• The custom NN is a promising tool for robust risk prediction in real-world medical datasets.