Radiomics Analysis Using Stability Selection Supervised Principal Component Analysis for Right-censored Survival Data

Kang K. Yan, Xiaofei Wang, Wendy Lam, Varut Vardhanabhuti, Anne W.m. Lee, Herbert PangbioRxiv 2018 · DOI

TL;DR

This paper introduces SSSuperPCA, a novel machine learning framework that combines stability selection with supervised principal component analysis to identify stable, prognostic radiomic features from high-dimensional medical images for predicting patient survival. The method demonstrates improved performance over existing techniques in predicting right-censored survival outcomes in non-small cell lung cancer and head and neck cancer datasets.

What Problem Does This Paper Address?

Predicting patient prognosis and therapeutic response is crucial for personalized cancer treatment. Radiomics, the extraction of quantitative features from medical images, offers a non-invasive way to glean this information. However, radiomic datasets are characterized by a massive number of features (high dimensionality) and often involve right-censored survival data (where the exact survival time is unknown for some patients). Traditional statistical and machine learning models struggle with this combination, leading to potential issues like overfitting and unreliable predictions. Existing radiomics studies often rely on generic algorithms not specifically tailored for this type of data, highlighting a need for specialized, robust methods.

Methods

The core innovation is the Stability Selection Supervised Principal Component Analysis (SSSuperPCA) algorithm. It's a two-stage process:

  1. Stability Selection with Boosted C-Index: This stage tackles the high dimensionality. It uses a boosted C-index (a measure of prediction accuracy for survival data) within a stability selection framework. Stability selection iteratively resamples data and selects features that are consistently identified across these resamples, thus identifying "stable" features. The boosted C-index helps optimize feature selection for survival prediction. This process is applied to subgroups of features ranked by their initial univariate prognostic ability to ensure a comprehensive selection.
  2. Supervised Principal Component Analysis (Supervised PCA): Once stable features are identified, Supervised PCA is applied. This technique uses these selected features to perform dimensionality reduction by creating principal components that are maximally correlated with the survival outcome. This helps to capture the underlying structure of the data while focusing on features relevant to survival, further refining the prediction model.

The paper also compares SSSuperPCA against several other machine learning algorithms, including supervised PCA (without stability selection), Cox proportional hazards regression, LASSO/Elastic-Net Cox regression, and Random Forests.

Datasets

The study utilizes two publicly available radiomics datasets:

  • Non-Small Cell Lung Cancer (NSCLC): This dataset consists of 316 patients with 208 events and a median survival of 543 days.
  • Head and Neck Cancer: This dataset comprises 79 patients with 15 death events.

Texture features were extracted from CT/RTSTRUCT images using Matlab, resulting in a large number of features (over 10,000 per dataset). The authors also conducted simulation studies using the NSCLC data to evaluate performance under various scenarios of sample size and the number of stable features.

Key Insights and Ideas

  • The Power of Stability: Stability selection is crucial for identifying truly prognostic features in high-dimensional radiomics data, effectively controlling for false positives and ensuring feature reliability.
  • Tailoring to Survival Data: Directly addressing right-censored survival data with specialized metrics like the C-index (and its truncated version) is essential for accurate prognostic modeling in cancer.
  • Combining Strengths: The synergy between stability selection (for robust feature identification) and supervised PCA (for dimensionality reduction focused on the outcome) creates a powerful framework for radiomics survival analysis.
  • Biologically Meaningful Features: The identified stable features often have clear biological interpretations (e.g., tumor shape, texture homogeneity), suggesting that the model is not just finding statistical correlations but also uncovering clinically relevant biomarkers.

Key Results

  • Improved Prediction Accuracy: SSSuperPCA consistently outperformed supervised PCA without stability selection and other benchmark algorithms across both real-world datasets. For NSCLC, SSSuperPCA increased the Uno's C-index from 0.588 to 0.628. For head and neck cancer, it achieved a C-index of 0.705, outperforming supervised PCA (0.658).
  • Simulation Study Success: In simulations, increasing the number of stable features and sample size generally improved SSSuperPCA's performance, with C-indices ranging from 0.550 to 0.624.
  • Top Features Identified: For NSCLC, the top stable features included tumor volume, size, and small zone emphasis. For head and neck cancer, features related to zone-size non-uniformity and gray-level emphasis were identified. These features are related to tumor shape and texture heterogeneity.

Limitations and Future Work

  • Cox Model Assumption: The paper acknowledges that the reliance on the Cox proportional hazards model in both stability selection and supervised PCA can be a limitation if the proportional hazards assumption is violated. The authors suggest exploring alternative models like the accelerated failure time model as a potential future direction.
  • Generalizability: While the approach is applied to CT scans, the authors believe it can be extended to other imaging modalities like MRI.
  • Further Algorithm Development: The authors hope their work will inspire further research into novel algorithms for bioimage informatics analysis, particularly for complex survival prediction tasks.

Summary updated

Abstract

Abstract Radiomics is a newly emerging field that involves the extraction of a large number of quantitative features from biomedical images through the use of data-characterization algorithms. Radiomics provides a noninvasive approach for personalized therapy decision by identifying distinctive imaging features for predicting prognosis and therapeutic response. So far, many of the published radiomics studies utilize existing out of the box algorithms to identify the prognostic markers from biomedical images that are not specific to radiomics data. T o better utilize biomedical image, we propose a novel machine learning approach, stability selection supervised principal component analysis (SSSuperPCA) that identify a set of stable features from radiomics big data coupled with dimension reduction for right censored survival outcomes. In this paper, we describe stability selection supervised principal component analysis for radiomics data with right-censored survival outcomes. The proposed approach allows us to identify a set of stable features that are highly associated with the survival outcomes, control the per-family error rate, and predict the survival in a simple yet meaningful manner. We evaluate the performance of SSSuperPCA using simulations and real data sets for non-small cell lung cancer and head and neck cancer, and compare it with other machine learning algorithms. The results demonstrate that our method has a competitive edge over other existing methods in identifying the prognostic markers from biomedical big imaging data for the prediction of right-censored survival outcomes. An R package SSSuperPCA is available at the website: http://web.hku.hk/∼herbpang/SSSuperPCA.html
BibTeX
@misc{yan2018radiomics,
  author = {Kang K. Yan and Xiaofei Wang and Wendy Lam and Varut Vardhanabhuti and Anne W.M. Lee and Herbert Pang},
  title = {{Radiomics Analysis Using Stability Selection Supervised Principal Component Analysis for Right-censored Survival Data}},
  year = {2018},
  doi = {10.1101/408831},
  howpublished = {bioRxiv},
}
Figure