Content
Summary
Synthetic simulations generate high-dimensional genomic data using Weibull-distributed time-to-failure models to evaluate binary classification and survival analysis methods. The framework creates censored datasets with specific correlation structures and class-dependent distributions, allowing for the comparison of algorithms under both proportional and violated hazard assumptions.
Intuition
To understand how we test survival analysis algorithms, imagine creating a digital patient population where we control every variable. We generate gene expression data where some genes (markers) are linked in blocks and others are noise. We then assign a 'time-to-failure' to each digital patient based on their gene profile using a Weibull distribution. Crucially, we introduce 'censoring' by setting a follow-up time; if a patient's failure time exceeds this limit, we only know they survived past that point, not exactly when they failed. This creates 'observed' classes that are slightly polluted with misclassified points, mimicking the messy reality of clinical data where we cannot always observe the true event. This setup lets us rigorously test if algorithms like TESA or BCSA can distinguish signal from noise and handle assumption violations.
Key equations
Scale parameter of the Weibull distribution, where is intercept, are coefficients drawn from exponential distribution, is the set of markers, and is the expression value of the ith marker.
Observed low-risk class , containing points from both true low-risk () and true high-risk () classes that have failure times greater than follow-up time .
Observed high-risk class S1, containing points from both true low-risk (Z0) and true high-risk (Z1) classes that have failure times less than follow-up time .
Derivation
- Generate features categorized as markers or non-markers, where markers form blocks of size k with dependent features within a block and independent features across blocks.
- Classify markers as global (homogeneous within each class) or heterogeneous (constructing subclasses where only one subclass differs from the reference class).
- Assign time-to-failure t in class y following a Weibull distribution with shape parameter κy (dependent only on class label) and scale parameter λ(z) (dependent on covariates).
- Calculate scale parameter λ(z) using the exponential linear combination of marker expression values.
- Assign points to observed classes S0 or S1 based on whether their failure time is greater than or less than the follow-up time .
- Recognize that observed classes S0 and S1 are 'polluted' with a small subpopulation of points from the opposite true class due to the stochastic nature of time-to-failure relative to Tf.
Worked example
Consider a simulation setting with 100 markers (20 global, 80 heterogeneous), block size k=5, and c=2 subclasses. If κ0 = 0.5 and κ1 = 1.5, does this setup satisfy Cox proportional hazards assumptions? Explain the impact on the observed classes.
- Identify the shape parameters: κ0 = 0.5 and κ1 = 1.5.
- Compare the shape parameters: Since κ0 ≠ κ1, the survivorship functions differ in shape.
- Determine assumption status: Violated Cox proportional hazards assumptions.
- Analyze observed classes: Points are assigned to S0 or S1 based on Tf. Because κ differs, the rate of failure over time differs between classes, affecting the proportion of points in S0 and S1 relative to the true classes Z0 and Z1.
Answer The assumptions are violated because κ0 ≠ κ1. This creates different survivorship functions, impacting the distribution of points in the observed classes S0 and S1 relative to the true classes Z0 and Z1.
Common mistakes
- Confusing the true risk classes (, ) with the observed risk classes (, ); and are derived from and based on censoring thresholds and may contain misclassified points.
- Assuming that the correlation parameters ρ0 and ρ1 must be equal; the extended model allows them to differ, which can impact the simulation results.
- Thinking that the shape parameter κ depends on covariates; the text specifies that κy depends only on the class label, while the scale parameter λ depends on covariates.