Independent Testing of Published CT Models for PD-L1 Status in Non-Small Cell Lung Cancer.

Area of Science:

• Radiology
• Oncology
• Medical Imaging

Background:

• The rise of perioperative programmed cell death protein-1 (PD-1) and PD-L1 immunotherapy in non-small cell lung cancer (NSCLC) necessitates preoperative assessment of PD-L1 status.
• Current methods for identifying immunotherapy candidates using CT-based features are limited by a lack of independent validation.

Purpose of the Study:

• To assess the performance of existing CT radiomic models in predicting PD-L1 status in an external, multi-institutional cohort of NSCLC patients undergoing surgery.

Main Methods:

• A literature review identified CT radiomic models for PD-L1 prediction (Jan 2017-Jul 2023).
• Three models were reconstructed using published features and coefficients.
• Model performance was evaluated on an external test set (n=225) using area under the receiver operating characteristic curve (AUC) analysis for PD-L1 tumor proportion score (TPS) thresholds (≥1% and ≥50%).

Main Results:

• Only 3 out of 17 identified models (18%) were reconstructible.
• Model 3 showed comparable discrimination for TPS≥50% (AUC, 0.61) in the external test set.
• Models 1 and 2 demonstrated significantly lower predictive performance compared to their originally published results for TPS≥50% and TPS≥1% respectively.
• A CD274 mRNA-fitted model (model 3a) showed moderate correlation with PD-L1 TPS and discriminated both thresholds (AUCs 0.61-0.66).

Conclusions:

• CT radiomic models can predict PD-L1 expression in resectable NSCLC at clinically relevant thresholds.
• Independent validation revealed that the predictive performance of these models is generally lower than initially reported in original publications.