Reproducibility of the radiosensitivity index and failure of CT radiomics as its surrogate: a public-data study in non-small cell lung cancer

  1. Zhe Chen
  2. Rei Gou
Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Purpose

The radiosensitivity index (RSI) and genomic-adjusted radiation dose (GARD) are increasingly treated as quantitative inputs to radiotherapy dose calculation. Two reproducibility issues bear on this use: whether CT radiomics can non-invasively recover RSI, and whether one coefficient of the equation is uniquely specified (printed CDK1 but implemented as PAK2). We examine both on public data.

Methods

In GEO GSE103584 RNA-seq (n = 130 non-small cell lung cancer [NSCLC]), we recomputed RSI with the Eschrich 2009 coefficients using CDK1 or PAK2 in the disputed slot and derived GARD under four fixed dose/fractionation schemas. In the paired TCIA NSCLC-Radiogenomics cohort (n = 117), we trained cross-validated Elastic Net and Random Forest models to predict continuous RSI and a median-split RSI label from IBSI-conformant, scanner-corrected CT radiomic features, under a pre-set viability rule.

Results

CT radiomics did not recover RSI (Spearman ρ = 0.05 and 0.03; binary AUC = 0.43), below the pre-set viability threshold. Separately, the two probesets listed for the disputed coefficient in the founding paper’s Table 3 both map to PAK2; using the printed CDK1 left rank correlation high (ρ = 0.980) but reclassified 6.2% and 9.2% of patients (median and tertile) and shifted GARD by 4.1-6.3 Gy.

Conclusions

CT radiomics is not a viable RSI surrogate in this public cohort, so imaging-GARD should not assume radiomic recovery of RSI. The disputed coefficient resolves to PAK2; implementing the printed CDK1 shifts GARD and reclassifies patients despite high rank correlation. Outcome-directed, dose-adjusted imaging is the more defensible next step.

Highlights

  • In 117 paired NSCLC cases, CT radiomics did not recover RSI (ρ≤0.05; AUC 0.43).

  • Imaging-GARD assuming a radiomic RSI surrogate is unsupported on public data.

  • RSI/GARD enter dose decisions, so each model term must be uniquely specified.

  • One RSI coefficient is printed inconsistently (CDK1 vs PAK2); probes support PAK2.

  • Using CDK1 not PAK2 shifts GARD 4-6 Gy and reclassifies 6-9% of patients.

Article activity feed

  1. Version published to 10.64898/2026.07.20.26358526 on medRxiv