Neoadjuvant immunotherapy has demonstrated promising efficacy in resectable lung cancer; however, the treatment response remains highly heterogeneous, and a substantial proportion of patients derive limited clinical benefits. Therefore, identifying patients who are most likely to benefit from immunotherapy before treatment initiation is essential for optimizing individualized treatment strategies and improving clinical outcomes.

In this study, clinical stage II, absence of lymph node metastasis, receipt of immunotherapy combined with chemotherapy, higher PD-L1 expression, and higher TMB were independently associated with an increased likelihood of achieving pCR after neoadjuvant immunotherapy. These findings are consistent with the current understanding of tumor biology and treatment response. Patients with earlier-stage disease generally have a lower tumor burden and a less immunosuppressive tumor microenvironment, which may facilitate immune cell infiltration and enhance responsiveness to immunotherapy [17]. Similarly, the absence of lymph node metastasis may reflect a more localized disease and preserved antitumor immune function, whereas nodal involvement is associated with systemic dissemination and immune suppression, potentially reducing treatment efficacy [18].

Our findings are broadly consistent with those of previous studies evaluating biomarkers and predictive models in the neoadjuvant immunotherapy setting for lung cancer. PD-L1 expression and TMB have been widely recognized as important predictors of response to immune checkpoint inhibitors; however, their predictive performance when used individually has been inconsistent across studies [19,20,21]. Consistent with these observations, both biomarkers were independently associated with pCR in our study, supporting their continued clinical relevance and highlighting the limitations of relying on a single biomarker. More recently, multivariable prediction models integrating clinical, molecular, and immunological characteristics have demonstrated improved predictive performance compared with individual biomarkers, although differences in study populations, variable selection, and modeling approaches have limited their generalizability [22, 23]. By integrating established clinical factors with key immunotherapy-related biomarkers and directly comparing conventional regression and machine learning approaches, our study further supports the value of multidimensional prediction models in the neoadjuvant setting.

The observed benefit of combining immunotherapy and chemotherapy may be explained by the synergistic effects of these treatment modalities. Cytotoxic chemotherapy reduces tumor burden and modulates the tumor microenvironment, thereby enhancing tumor antigen release and immune recognition, whereas immune checkpoint inhibition restores T cell-mediated antitumor activity [24]. Consequently, combined therapy may increase the likelihood of achieving pCR compared with immunotherapy alone.

PD-L1 expression and TMB were also identified as the two most influential variables in the random forest model. PD-L1 expression reflects tumor immune evasion through activation of the PD-1/PD-L1 pathway, whereas TMB serves as a surrogate marker of tumor immunogenicity by reflecting the neoantigen burden [25, 26]. Integrating these complementary biomarkers may provide a more comprehensive assessment of the tumor immune microenvironment and improve the prediction of treatment response. Therefore, our findings support the potential value of combining PD-L1 expression and TMB in biomarker-guided treatment strategies for patients undergoing neoadjuvant immunotherapy [27].

The RF model demonstrated superior discrimination, specificity, and overall classification performance compared with the nomogram, while maintaining acceptable performance in both the internal and temporally independent validation cohorts. In contrast, the nomogram demonstrated higher sensitivity, suggesting that it may be more suitable for identifying potential responders, whereas the random forest model may better reduce false-positive predictions. These findings highlight the complementary strengths of traditional regression-based models and machine learning approaches for clinical predictions.

From a clinical perspective, the proposed model may support biomarker-informed risk stratification by identifying patients with a higher probability of achieving pCR, thereby facilitating individualized neoadjuvant treatment strategies. Such predictive tools may also assist multidisciplinary teams in integrating molecular biomarker information into therapeutic decision-making.

The potential clinical utility of the proposed model warrants consideration. By providing individualized estimates of the probability of achieving pCR, these models may facilitate risk stratification and support personalized treatment planning for patients undergoing neoadjuvant immunotherapy. Patients predicted to have a high probability of achieving pCR may be appropriate candidates for standard immunotherapy-based regimens, whereas those predicted to have a lower probability may benefit from alternative treatment strategies or enrollment in clinical trials. Consequently, these models may help reduce unnecessary exposure to ineffective treatments while supporting more individualized care. However, their clinical utility should be confirmed in prospective studies with larger sample sizes. Future investigations incorporating decision curve analysis and prospective multicenter validation are warranted to evaluate the net clinical benefits and real-world applicability of these models.

This study has several limitations. First, this was a retrospective single-center study, and selection bias and unmeasured confounding factors cannot be excluded. Second, although temporal validation was performed, calibration in the temporally independent validation cohort was suboptimal. This may reflect differences in patient characteristics, treatment regimens, and biomarker distributions between cohorts. In particular, treatment heterogeneity, including differences in chemotherapy regimens, immune checkpoint inhibitors, and treatment schedules, may have contributed to variations in model performance. Consequently, although the model demonstrated acceptable discrimination, caution is warranted when interpreting the absolute risk estimates. Future studies should evaluate model updating and recalibration strategies, including the recalibration of the intercept and slope, using larger independent cohorts. Third, the relatively limited sample size may have affected model stability despite internal and temporal validation. Because of the retrospective design, no formal sample size calculations were performed. Fourth, PD-L1 assessment was based on a single assay platform, and TMB measurement methods and cut-off values may differ across institutions, potentially limiting generalizability. Fifth, additional biomarkers, including tumor-infiltrating lymphocytes, microsatellite instability status, and driver gene mutations, were not incorporated and may further improve predictive performance. Finally, this study focused on pCR as a short-term endpoint, and long-term outcomes, including disease-free survival and overall survival, were not evaluated. Prospective multicenter studies with longer follow-up periods are required to further validate and refine these prediction models.