Survival ensembles by the sum of pairwise differences with application to lung cancer microarray studies

Brent A. Johnson, Qi LongThe Annals of Applied Statistics 2011 · DOI

TL;DR

This paper introduces a novel "survival ensemble" method for analyzing time-to-event data, particularly useful in high-dimensional settings like cancer genomics. Their approach, based on an accelerated failure time (AFT) model and minimizing the sum of pairwise differences in residuals, offers a more robust alternative to traditional proportional hazards (PH) models, especially when PH assumptions are violated.

What Problem Does This Paper Address?

Analyzing survival data, where the outcome is the time until an event occurs (e.g., death, disease recurrence), is crucial in many fields, especially medicine and biology. A common challenge is dealing with "high-dimensional" data, where the number of potential predictors (like genes in a microarray study) is much larger than the number of observations. Traditional methods often rely on the proportional hazards (PH) assumption, which states that the hazard rate (instantaneous risk of the event) is constant over time for individuals with different predictor values. However, this assumption can be violated in real-world data, leading to inaccurate conclusions, particularly in variable selection. This paper aims to develop a more flexible and robust statistical learning method for survival analysis that doesn't strictly adhere to the PH assumption, especially in high-dimensional scenarios.

Methods

The core of this paper is a new boosting algorithm for survival data. Boosting is a machine learning technique that builds a strong predictive model by sequentially combining many weak learners. Here, the authors adapt boosting to survival analysis by:

  • Accelerated Failure Time (AFT) Model: Instead of modeling the hazard rate (like PH models), they use the AFT model, which directly models the logarithm of the survival time as a linear function of predictors. This model is more interpretable as it relates predictor changes to changes in survival time itself.
  • Sum of Pairwise Differences Loss Function (Gehan Loss): They propose a novel loss function based on minimizing the sum of pairwise differences in residuals. This "Gehan loss" is derived from rank-based tests and is less sensitive to the strict assumptions of the PH model. It essentially measures how well the model orders pairs of individuals based on their predicted survival times.
  • Rank-Based Boosting: They apply boosting to this AFT model with the Gehan loss function. This means they iteratively build an ensemble of simple models (like regression trees) that, when combined, minimize the Gehan loss. This approach is less reliant on the proportional hazards assumption.

They contrast this with boosting based on the proportional hazards model (using partial likelihood) and inverse-probability weighting (IPW) methods.

Datasets

The paper uses two primary datasets for evaluation:

  • Lung Cancer Microarray Data: This dataset, previously analyzed by Morris et al. (2005), contains gene expression data from lung adenocarcinoma patients. The goal is to identify genes associated with mortality. This dataset is characterized by a high number of genes (predictors) and a relatively small number of patients.
  • Nursing Home Data: This dataset, previously analyzed by Fan and Li (2002), investigates the effect of financial incentives on patient care in nursing homes. This dataset is characterized by a large sample size relative to the number of predictors, allowing for more standard diagnostic tests.

Key Insights and Ideas

  • PH Assumption is Fragile: The study strongly suggests that the proportional hazards assumption is often violated in real-world data, particularly in high-dimensional settings like genomics. Diagnostic tests on the microarray data indicated this violation.
  • AFT Models Offer Robustness: The proposed AFT-based approach with the Gehan loss function provides a valuable alternative when PH assumptions are not met. It can identify different sets of important predictors compared to PH models.
  • Boosting is Versatile: The paper demonstrates the power of boosting as a flexible machine learning framework that can be adapted to various survival models (PH and AFT) and loss functions.
  • Model Misspecification Matters: Simulation studies highlight that misspecifying the PH model can lead to underfitting, ignoring moderately sized effects, and biased variable selection, even when the underlying data-generating process is different.
  • Different Loss Functions Lead to Different Insights: The comparison of Gehan loss with other loss functions (like IPW) shows that the choice of loss function can significantly impact the identified predictors and model conclusions, even within the same AFT framework.

Key Results

  • Lung Cancer Microarray Analysis:
    • The proposed rank-based boosting method identified a 19-gene panel associated with lung cancer mortality.
  • A standard PH boosting approach identified only a 9-gene panel, which was a subset of the 19 genes found by the new method.
  • This discrepancy suggests that the PH assumption might have led to the exclusion of important genes.
  • Simulation Studies:
    • When the PH model was misspecified, PH boosting tended to underfit and ignore moderately sized covariate effects.
  • In scenarios where the PH model was incorrect, the AFT-based rank-boosted models showed less bias and better performance in identifying relevant variables compared to PH models.
  • The choice of loss function in boosting (e.g., Gehan vs. IPW) could lead to substantial differences in predictive scores and identified variables.

Limitations and Future Work

  • Computational Cost: Boosting, especially with complex loss functions and many iterations, can be computationally intensive, as noted by the authors.
  • Parameter Tuning: The performance of boosting is sensitive to parameter tuning (e.g., the number of iterations, mstop). While cross-validation is used, finding optimal parameters can still be challenging.
  • Diagnosing PH Assumption in High Dimensions: The paper acknowledges that diagnosing the PH assumption in high-dimensional settings is an "intriguing research problem" and beyond the scope of the current manuscript.
  • Future Work: The authors suggest that their rank-based survival ensembles offer a powerful tool for investigating large datasets and encourage further research into diagnosing PH model assumptions in high-dimensional data. They also highlight the need for robust methods that can handle various error distributions and censoring mechanisms.

Summary updated

Abstract

Lung cancer is among the most common cancers in the United States, in terms of incidence and mortality. In 2009, it is estimated that more than 150,000 deaths will result from lung cancer alone. Genetic information is an extremely valuable data source in characterizing the personal nature of cancer. Over the past several years, investigators have conducted numerous association studies where intensive genetic data is collected on relatively few patients compared to the numbers of gene predictors, with one scientific goal being to identify genetic features associated with cancer recurrence or survival. In this note, we propose high-dimensional survival analysis through a new application of boosting, a powerful tool in machine learning. Our approach is based on an accelerated lifetime model and minimizing the sum of pairwise differences in residuals. We apply our method to a recent microarray study of lung adenocarcinoma and find that our ensemble is composed of 19 genes while a proportional hazards (PH) ensemble is composed of nine genes, a proper subset of the 19-gene panel. In one of our simulation scenarios, we demonstrate that PH boosting in a misspecified model tends to underfit and ignore moderately-sized covariate effects, on average. Diagnostic analyses suggest that the PH assumption is not satisfied in the microarray data and may explain, in part, the discrepancy in the sets of active coefficients. Our simulation studies and comparative data analyses demonstrate how statistical learning by PH models alone is insufficient.
BibTeX
@article{johnson2011survival,
  author = {Brent A. Johnson and Qi Long},
  title = {{Survival ensembles by the sum of pairwise differences with application to lung cancer microarray studies}},
  journal = {The Annals of Applied Statistics},
  year = {2011},
  doi = {10.1214/10-aoas426},
}
Figure