Immunochemotherapy combining platinum–etoposide with anti–PD-L1 agents is a standard approach for extensive-stage small-cell lung cancer (SCLC), but reliable peripheral biomarkers to predict treatment response and prognosis are lacking. Prior univariate analyses have suggested associations between individual lymphocyte subsets and survival, yet the systemic immune response is inherently multivariate. This study used high-dimensional peripheral blood flow cytometry and a robust machine-learning pipeline to determine whether multivariate patterns in circulating immune cells can provide an early prognostic signal in patients receiving carboplatin, etoposide, and atezolizumab.
The analysis included 32 patients with stage IV SCLC treated with carboplatin, etoposide, and atezolizumab. Peripheral blood was collected at baseline (V0) and longitudinally during treatment; to prioritise early biomarkers and mitigate sample attrition in later treatment cycles, the primary comparisons focused on baseline (V0) and measurements after two cycles of immunochemotherapy (V1). The study report is a medRxiv preprint and has not been peer reviewed.
High-dimensional flow cytometry was performed on peripheral blood samples to quantify multiple lymphocyte subsets. The analysis targeted both absolute cell counts at specified visits and dynamic changes between visits (∆V, i.e., change from V0 to V1). Specific technical preprocessing steps are described in the full manuscript; the present summary reports the analytical strategy and the principal outcomes derived from the processed data.
To evaluate multivariate prognostic signals, the authors implemented a rigorous machine-learning framework. Eleven different regression and survival models were assessed. The evaluation used nested cross-validation to guard against overfitting, bootstrapping to estimate stability, and permutation-based statistical testing to assess significance under appropriate null distributions. Model-appropriate performance metrics were applied: standard regressors were evaluated with regression metrics, while survival models were evaluated with censoring-aware metrics such as the concordance index (C-index). Where relevant, permutation testing produced p values for observed performance relative to chance.
Under the applied evaluation framework, standard regressors failed to generalise, producing R2 values below zero under the chosen metrics. In contrast, censoring-aware Random Survival Forests (RSF) extracted prognostic signatures that were statistically significant.
These results emphasise the value of baseline and early dynamic multivariate signatures over single timepoint absolute counts after treatment initiation.
Within the RSF-derived prognostic signatures, feature-importance analysis identified components of the regulatory–effector immune balance. The analysis corroborated the prognostic relevance of Th17 normalisation and nominated Naive Regulatory T cells and Memory B cells as candidate contributors to the multivariate signature. The report presents these features as candidate biomarkers requiring further validation rather than definitive clinically validated markers.
The study provides a proof of concept that multivariate peripheral immune profiling, analysed with censoring-aware machine learning methods, can reveal early prognostic signals in extensive-stage SCLC treated with immunochemotherapy. The findings suggest that early dynamic shifts in the balance between regulatory and effector immune compartments are associated with outcome, while absolute counts after two cycles lack prognostic information in this cohort. The authors highlight the need to confirm these observations in larger cohorts and to integrate systemic immune data with tumour-intrinsic information to improve predictive accuracy.
Key limitations reported in the preprint include the small sample size (n = 32) and the preprint status of the manuscript (not peer reviewed). The authors explicitly call for confirmation in larger patient cohorts and for studies that combine peripheral immune profiling with tumour-intrinsic data. The report frames the presented results as exploratory and proof-of-concept rather than immediately practice-changing.
The study obtained ethical approval from the Institutional Review Board of the University Hospital Augsburg (protocol code 2020-10; approved 25 April 2020) and written informed consent was secured from participants. During data handling, identifying information was accessible to clinical authors affiliated with the University Hospital Augsburg and the data were pseudonymised before transfer to computational authors. Data produced in the study are available from the authors upon reasonable request. The manuscript is posted as a medRxiv preprint and has not been certified by peer review.