Explainable Machine Learning–Based Survival Analysis for Predicting Breast Cancer Recurrence

Breast cancer can come back. This is a big, hard task for doctors. Fast, right signs help them act & may help more people live. Usual ways to look at who might get sick again, like the Cox model, are not good with hard links in the data. These ways are also hard for doctors to trust or use, as they are not clear. We set out to make & check a clear ML way to look at who may get breast cancer again. We used many kinds of data, like health notes, tissue photos, & gene info. We found data from about ۲,۰۰۰ people. We split the group as ۷۰% to train, ۱۵% to check, & ۱۵% to test. The data was fixed to deal with missed info, mix signs from all parts, & pick the best ones to use. We did this with a way that looks for the top signs, made clear by SHAP. Our model used DeepSurv with parts that show which signs matter more. It also used SHAP to show points both for all & for each case. We looked at how well the model worked with the C-index, Brier score, & test lines by time. We also asked top doctors what they thought. The model got a C-index of ۰.۸۱. This is better than the Cox model, which got ۰.۷۵. The model did well for ۱-, ۳-, & ۵-year marks. The top signs that told the risk were things like how mean the lump is, if nodes were hit, HER۲, TP۵۳ changes, & how the cell’s core looks. Doctors said these signs made sense, which helped them trust the tool. This tells us that clear ML can have both high scores & be clear for use. It lets us split risks well & helps doctors trust what the tool shows. These new ways with good, clear ML can fit with doctor work & help make much better plans. In time, this may cut down on breast cancer coming back.

کلیدواژه ها:

Breast cancer recurrence prediction ، survival analysis ، explainable machine learning ، multimodal data integration ، clinical interpretability