Document Type: Research Articles
Niramai Health Analytix Pvt Ltd., Koramangala, Bangalore, Karnataka, India.
Department of Radiation Oncology (MAASTRO Clinic), GROW School for Oncology and Developmental Biology, Maastricht University Medical Centre+, Maastricht, The Netherlands.
Purpose: To evaluate the robustness of multiple machine learning classifiers for breast cancer risk estimation in the presence of incomplete or inaccurate information. Data and methods: Open data for this study was obtained from the BCSC Data Resource (http://breastscreening.cancer.gov/). We conducted two ablation-type experiments to compare the robustness of different classifiers where we randomly switched known information to missing with a missing probability of pm in one experiment, and randomly corrupted the existing information with a probability of pc in another experiment. We considered three prominent machine-learning classifiers such as Logistic regression (LR), Random Forests (RF) and a custom Neural Network (NN) architecture and compared their degradation of discrimination performance as a function of increasing probability of missing or inaccurate data. Results: LR, RF and custom NN resulted in an Area Under Curve (AUC) of 0.645, 0.643 and 0.649, respectively, on a test set with 500,000 total observations. When we manipulated the data by varying probabilities pm and pc from 0 to 1, NN resulted in better performance in terms of AUC compared to RF and LR as long as less than half the data was missing/inaccurate (that is, for values of pm < 0.5 and pc < 0.5). However, for missing (pm) or corruption (pc) probabilities above 0.5, LR gave similar performance as the custom NN. RF resulted in overall poorer performance when the data had additional missing or incorrect entries. Conclusion: In cases where the input information is missing or inaccurate, our experiments show that the proposed custom NN provides reliable risk estimates in medical datasets like BCSC. These results are particularly important in health care applications where not every attribute of the individual participant might be available.