A deep learning biomarker derived from routine serial computed tomography (CT) scans has been shown to predict overall survival (OS) more accurately than conventional response criteria in patients with advanced non-small cell lung cancer (NSCLC) treated with immunotherapy.

Reliable early indicators of long-term benefit from ICIs remains an unmet need in advanced NSCLC. Current imaging-based measures, including Response Evaluation Criteria in Solid Tumors (RECIST) and tumour volume change (TVC), are widely used but have limited predictive value for OS, particularly given the complexity of immunotherapy response patterns.

Research published in JAMA Network Open aimed to develop and externally validate a fully automated imaging-based biomarker – the Serial CT Response Score (Serial CTRS) – derived from baseline and early follow-up CT scans to predict OS in patients with advanced NSCLC treated with immune checkpoint inhibitors (ICIs).

Using retrospectively collected electronic health record data from routine clinical practice across 10 European and US institutions, together with clinical trial data from the multinational phase 1 GARNET trial of dostarlimab, the study analysed 1,830 adults with advanced NSCLC who were treated with ICI monotherapy or combination therapy between 2013 and 2023.

The routine clinical practice discovery dataset comprised 1,171 patients, with a further 605 patients in the routine practice test dataset. Independent external validation was performed in 54 patients enrolled in the GARNET trial.

The median age across cohorts was 67 years, with 55% male and 45% female. Baseline CT scans and follow-up scans at approximately 12 weeks after treatment initiation were analysed.

Outperforming RECIST and TVC

Serial CTRS was independently associated with NSCLC OS across datasets after adjustment for age, sex, histological subtype, programmed death-ligand 1 (PD-L1) expression and baseline tumour volume.

In the routine practice test dataset, each 10-percentage-point increase in predicted 12-month survival probability was associated with a 27% reduction in the hazard of death (hazard ratio [HR] 0.74; 95% CI 0.70–0.79). In the GARNET validation cohort, the corresponding HR was 0.45 (95% CI 0.32–0.65).

Serial CTRS consistently outperformed RECIST and TVC in discriminating survival risk. In the routine practice test cohort, the HR comparing low versus high survival probability groups was 6.19 (95% CI 4.12–9.28) for Serial CTRS, compared with 3.79 for RECIST and 4.55 for TVC.

Similar findings were observed in the GARNET cohort, where the HR for low versus high survival probability reached 18.00 (95% CI 5.40–59.97).

Importantly, the biomarker retained prognostic value across PD-L1 expression subgroups and in patients with stable NSCLC by RECIST at 12 weeks.

Clinical implications for immunotherapy in NSCLC

Relatively high censoring rates in the real-world development dataset and the limited interpretability inherent to deep learning models were flagged as constraints.

Serial CTRS analyses were confined to thoracic imaging, which might not capture prognostic information from other metastatic sites, and the authors further noted that a small sample size in the clinical trial validation cohort resulted in wide confidence intervals for some estimates.

However, the findings suggest that Serial CTRS could provide an objective, early indicator of survival benefit using routinely acquired CT scans, potentially supporting treatment decisions in patients with equivocal early responses to immunotherapy for NSCLC.

The authors concluded that further validation across additional tumour types and treatment modalities was warranted, and that integrating Serial CTRS into clinical trials could potentially improve early assessment of therapeutic benefit in clinical and research settings.

Reference
Sako C et al. Deep-Learning Serial CT Prediction of Survival in Immunotherapy-Treated Non-Small Cell Lung Cancer. JAMA Netw Open 2026;9(1):e2555759.