Combining Clinical Data and Longitudinal Neutrophil Profiling for Survival Prediction in Immunotherapy-treated aNSCLC: an Explainable Machine Learning Analysis

Most patients with advanced non-small-cell lung cancer receive immunotherapy as first-line treatment, alone or combined with chemotherapy. However, survival after treatment varies among patients, and identifying early who is at higher risk of relapse would allow clinicians to adjust therapy in time. Circulating low-density neutrophils (LDNs) are a plastic immune population whose baseline levels have been associated with treatment response [1]. However, neutrophils comprise subpopulations that span a continuum of maturation states, and their composition may change during treatment. In this study, we applied explainable machine learning to assess the value of profiling baseline and on-treatment LDN subpopulations, leveraging the availability of longitudinal data, which is rare in this setting. We focus on predicting overall survival (OS) and progression-free survival (PFS), two common clinical outcomes related to treatment efficacy. We built different predictive models using two cohorts from an Italian cancer center, with patients receiving immunotherapy (IO, n = 118) or chemo-immunotherapy (IO-CT, n = 179), because the mechanisms underlying these treatments differ significantly. Blood was sampled by flow cytometry [2] before treatment start (T0), after one cycle (T1), and at the first radiological evaluation (T2). This produced 8 LDN subpopulations per timepoint. 15 baseline clinical variables were included, i.e., age, sex, BMI, ECOG PS, PD-L1, LDH, NLR, and metastatic sites. Three different sets of features were considered for training: (i) using clinical data and LDNs at T0; (ii) adding LDNs at T1; and (iii) adding LDNs at T2. Since not all patients had all time points available, we opted to use XGBoost for its ability to handle missing data. We evaluate the models by means of Harrell’s C index, computed using repeated (10 runs) nested cross-validation (we report mean values ± standard deviations). Finally, we used SHapley Additive exPlanations (SHAP) for model explainability. The model using baseline data (T0) can already discriminate OS, with a C-index of 0.68 ± 0.02 in the IO cohort and 0.69 ± 0.02 in the IO-CT cohort, and PFS, with C-indices of 0.66 ± 0.02 and 0.63 ± 0.02, respectively. Adding T1 and T2 data left performance unchanged, with only a marginal increase in OS in the IO cohort, i.e., 0.69 ± 0.02 with T1 features and 0.71 ± 0.02 with T1 and T2. The models for the IO cohort across different time points relied on different features. Indeed, the SHAP plots revealed that at baseline (T0), risk was dominated by the patient’s starting profile, although the composition of that profile differed between cohorts. In IO, the model prediction was driven by total myeloid abundance (i.e., CD11b+ cells), with higher values indicating higher risk, whereas in IO-CT it was driven by clinical variables (i.e., ECOG PS, NLR, and BMI). In both cohorts, a higher fraction of mature CD10+ cells was associated with lower risk. When adding T1, baseline features remained the most important, while the highest-ranked T1 features were the mature CD10+ fraction in IO and total myeloid abundance in IO-CT, in the same directions as at baseline. As shown in Figure 1, when adding T2, myeloid abundance and a lower mature CD10+ fraction at T2 were associated with a higher risk for OS in IO. Taken together, different biological mechanisms appear to happen in the two cohorts. Within the IO setting, the myeloid compartment already carries prognostic information, and its failure to mature during treatment further augments its prognostic value. Under IO-CT, risk remains driven by the clinical profile, consistent with the chemotherapy-induced neutrophil depletion. Future validations may enable adaptive strategies for high-risk patients through low-cost and minimally invasive immune monitoring. [1] N Castro et al. “Circulating low-density neutrophils as biomarkers of resistance to first-line anti-PD-1/PDL1 immunotherapy in non-small cell lung cancer”. In: Translational Oncology 67 (2026), p. 102755. [2] Orla Maguire et al. “Guidelines for the use of flow cytometry and cell sorting in immunological studies”. In: European journal of immunology 49.10 (2019), pp. 1457–1973