Content
Summary
Binary Classification for Survival Analysis (BCSA) categorizes subjects into discrete risk groups, offering an alternative to traditional Time to Event Survival Analysis (TESA). To handle random censoring, BCSA excludes censored points during parameter estimation but applies the resulting model to them afterward. This approach is particularly useful when censoring is minimal, markers are weak, or TESA assumptions are violated.
Intuition
Think of TESA as trying to predict the exact minute a machine breaks, while BCSA simply asks if it breaks before a specific deadline. Because BCSA relies on knowing the final outcome (broken or not) to learn the rules, it cannot use data points where the outcome is unknown due to censoring during the training phase. However, once the rule is set, it can guess the status for those censored items. This method shines when the signals are weak or the standard assumptions of survival models fail, provided the amount of missing data is small.
Derivation
- BCSA differs from TESA by outputting discrete classes rather than continuous hazard functions.
- Randomly censored observations have unknown event status, making them unsuitable for training a classifier.
- Therefore, censored points must be excluded from the training phase to estimate classifier parameters without bias.
- After the classification rule is established using non-censored data, the model is applied to the excluded censored points.
- This two-step process allows for the classification of the entire cohort post-modeling.
Worked example
A researcher has a dataset with 100 subjects, 20 of which are randomly censored. They wish to use BCSA to predict failure within 5 years. How should they handle the 20 censored subjects during the initial model building?
- Identify that BCSA requires known event status for training.
- Recognize that censored subjects have unknown status.
- Exclude the 20 censored subjects from the training dataset.
- Train the classifier on the remaining 80 non-censored subjects.
- Apply the trained model to predict labels for the 20 censored subjects.
Answer Exclude the 20 censored subjects from training, build the model on the 80 known cases, and then predict the labels for the 20 censored subjects using the trained model.
Common mistakes
- Assuming BCSA can directly handle random censoring by assigning censored points to classes during training.
- Conflating classification-based feature selection with False Discovery Rate (FDR) correction.
- Ignoring that using many weak covariates increases False Discovery Proportion (FDP), hindering biological interpretation.