[Paper Review] Stable prediction with radiomics data

Carel F.W. Peeters, Caroline Übelhör+10 more|arXiv (Cornell University)|Mar 27, 2019

Radiomics and Machine Learning in Medical ImagingMedicine34 references3 citations

TL;DR

This paper proposes FMradio, a four-step method that stabilizes radiomic prediction by projecting high-dimensional, multicollinear features into a lower-dimensional latent space via redundancy-filtered factor analysis. It outperforms existing methods in survival prediction for squamous cell cancers, achieving the lowest prediction error and highest explained variation in both internal and external validation.

ABSTRACT

Motivation: Radiomics refers to the high-throughput mining of quantitative features from radiographic images. It is a promising field in that it may provide a non-invasive solution for screening and classification. Standard machine learning classification and feature selection techniques, however, tend to display inferior performance in terms of (the stability of) predictive performance. This is due to the heavy multicollinearity present in radiomic data. We set out to provide an easy-to-use approach that deals with this problem. Results: We developed a four-step approach that projects the original high-dimensional feature space onto a lower-dimensional latent-feature space, while retaining most of the covariation in the data. It consists of (i) penalized maximum likelihood estimation of a redundancy filtered correlation matrix. The resulting matrix (ii) is the input for a maximum likelihood factor analysis procedure. This two-stage maximum-likelihood approach can be used to (iii) produce a compact set of stable features that (iv) can be directly used in any (regression-based) classifier or predictor. It outperforms other classification (and feature selection) techniques in both external and internal validation settings regarding survival in squamous cell cancers.

Motivation & Objective

• To address the instability of standard machine learning classifiers in radiomic data due to high-dimensional multicollinearity.
• To develop a robust, easy-to-use pipeline for downstream analysis of radiomic features in classification and prediction tasks.
• To improve predictive performance and feature stability in survival modeling for squamous cell cancers using radiomic data.
• To provide a dedicated, validated workflow for radiomic downstream analysis, filling a gap in current radiomics research.

Proposed method

• First, apply penalized maximum likelihood estimation to a redundancy-filtered correlation matrix to reduce multicollinearity.
• Second, perform maximum likelihood factor analysis on the filtered correlation matrix to extract latent factors.
• Third, use the resulting factors as a compact, stable set of features for downstream modeling.
• Fourth, directly integrate the stable latent features into any regression-based classifier or predictor.
• The method combines redundancy filtering with two-stage maximum-likelihood estimation to ensure stability and interpretability.
• The FMradio R package implements the full pipeline, enabling reproducible and scalable analysis.

Experimental results

Research questions

• RQ1 Can a two-stage maximum-likelihood approach reduce multicollinearity in high-dimensional radiomic data and improve prediction stability?
• RQ2 How does FMradio compare to standard feature selection and classification techniques in terms of predictive performance and stability?
• RQ3 Does redundancy filtering enhance the performance of survival models in radiomic data?
• RQ4 Can the latent features derived from factor analysis serve as robust predictors in external validation settings?
• RQ5 What is the impact of multicollinearity on the uniqueness and stability of feature selection in radiomic classification?

Key findings

• FMradio achieved the lowest cross-validated integrated Brier score of 0.108, significantly outperforming other methods including Cox boosting (0.109) and random survival forests (0.113).
• The method achieved an averaged cross-validated explained variation (R²) of 0.169, the highest among all evaluated approaches.
• In internal validation, FMradio showed the most stable prediction error curves and highest R² values, indicating superior robustness.
• Redundancy filtering improved performance across all competing methods, but FMradio remained the most stable and accurate overall.
• The latent features derived from factor analysis were consistently predictive and stable across both internal and external validation settings.