Explainable AI reveals squamous histology and U-shaped PD-L1 patterns as primary subgroup predictors of neoadjuvant and perioperative immunotherapy benefit in NSCLC: a machine learning analysis

Abstract

Background

The introduction of neoadjuvant and perioperative immunotherapy has broadened treatment options for resectable non-small cell lung cancer (NSCLC). However, clinical benefit varies across subpopulations, and standard linear models cannot fully capture the complex feature interactions and trial-level differences found in aggregate data.

Methods

We applied an integrated framework combining multilevel meta-regression with an inverse-variance weighted eXtreme Gradient Boosting (XGBoost) algorithm. SHapley Additive exPlanations (SHAP) were used to interpret the model and identify efficacy variance drivers across seven randomized controlled trials.

Results

Multilevel meta-regression demonstrated that event-free survival (EFS) benefit correlated positively with PD-L1 expression, peaking in the $\ge$50% subgroup (adjusted HR = 0.44, 95% CI: 0.33-−0.57). XGBoost-SHAP analysis revealed trial-level variance as the dominant driver of heterogeneity. Among subgroup-level clinical covariates, a non-linear U-shaped PD-L1 predictive pattern (<1% and $\ge$50%), squamous cell carcinoma (SCC) histology (HR = 0.53, 95% CI: 0.43-−0.65), and smoking history emerged as primary predictors. Synthesizing these signatures into a hypothesis-generating subgroup stratification framework indicated that SCC and PD-L1-negative (<1%) non-squamous cohorts benefit from continuous perioperative blockade. Conversely, exploratory analyses suggested that PD-L1-positive ($\ge$1%) non-squamous tumors achieved maximal observed benefit from exclusively neoadjuvant regimens (HR = 0.50).

Conclusions

Our results suggest that for PD-L1-positive non-squamous cases, the added benefit of extended adjuvant therapy may be limited. However, given our reliance on aggregate data and trial-level imbalances, these findings remain hypothesis-generating and should not alter current clinical practice. Rather, they offer an exploratory framework to inform patient selection for future de-escalation trials.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00262-026-04450-8.

Background

The integration of neoadjuvant and perioperative immune checkpoint inhibitors (ICIs) has transformed the management of resectable non-small cell lung cancer (NSCLC) [1, 2]. Following pivotal phase III trials–including CheckMate 816 [3], KEYNOTE-671 [4], AEGEAN [5], Neotorch [6], CheckMate 77T [7], and RATIONALE-315 [8]–ICI-based regimens constitute the established standard of care. However, the rapid succession of overlapping treatment approvals has shifted the clinical challenge from a lack of therapeutic options to the complexity of patient selection and regimen sequencing [2]. Specifically, the routine application of continuous, year-long perioperative ICI therapy raises critical concerns regarding overtreatment. It remains unclear which subpopulations require adjuvant consolidation compared to those achieving equivalent outcomes with a purely neoadjuvant approach [9]. Minimizing cumulative toxicity–including immune-related adverse events (irAEs) and financial burden–without compromising survival demands the precise personalization of therapeutic duration [10]. Consequently, identifying reliable clinical predictors to guide treatment de-escalation is a major priority in thoracic oncology [11].

Current evaluations of subpopulation efficacy rely on univariable subgroup analyses and conventional meta-analyses [12]. These frameworks assess clinico-pathological features (e.g., histology, sex) and molecular biomarkers (e.g., PD-L1 expression) in isolation. Because they assume clinical features act independently, these methods often overlook the high-dimensional interactions defining the tumor-immune microenvironment [13]. Aggregate-data (AD) meta-analyses are also inherently susceptible to aggregation bias (the ecological fallacy). Such approaches have limited capacity to adjust for trial-level confounding driven by heterogeneous study architectures, including differing chemotherapy backbones and varying immune checkpoint blockade durations [14, 15]. Relying on univariable aggregate estimates can therefore yield inconsistent predictive signals, complicating the development of precise, response-adapted treatment guidelines.

Integrating machine learning (ML) with conventional meta-analytical statistics provides an approach to address the limitations of univariable summary estimates. While multilevel meta-regression accounts for trial-level structural imbalances, tree-based ensemble algorithms–specifically, inverse-variance weighted eXtreme Gradient Boosting (XGBoost) [16]–can evaluate multifactorial associations across diverse clinical strata. Although applying ML to aggregate data cannot reconstruct individual-level biological interactions and carries an inherent risk of overfitting, these constraints can be managed statistically. By employing stringent regularization (e.g., restricting maximum tree depth) to control model complexity and integrating subgroup-specific inverse-variance weights, the algorithm prioritizes clinical subgroups with narrow confidence intervals (greater statistical precision) while down-weighting highly variable, small-sample estimates. This strategy effectively mirrors the conservative nature of a random-effects meta-analysis within a non-linear computational space. When interpreted via SHapley Additive exPlanations (SHAP) [17], the framework avoids spurious causal claims. Instead, it yields a transparent ranking of clinical predictors to generate data-driven hypotheses for patient stratification–an analytical paradigm increasingly validated in precision oncology [18–20].

We synthesized subgroup-level event-free survival (EFS) estimates from seven pivotal NSCLC trials. Our approach integrated a multilevel meta-regression model–incorporating empirical Bayes shrinkage to adjust for baseline imbalances and trial-level heterogeneity [21]–with an inverse-variance weighted machine learning pipeline to evaluate non-linear predictors of treatment effect heterogeneity. Based on the identified subgroup-level clinical predictors, we propose a hypothesis-generating stratification framework driven by histology and PD-L1 status. This provides data-driven insights to inform the design of future perioperative regimens and help prioritize candidates for treatment de-escalation trials [10].

Methods

We extracted aggregate survival data from seven pivotal phase II and III randomized controlled trials evaluating perioperative and neoadjuvant immune checkpoint inhibitors (ICIs) in resectable non-small cell lung cancer (NSCLC): AEGEAN [5], CheckMate 816 [3], CheckMate 77T [7], KEYNOTE-671 [4], NADIM II [22], Neotorch [6], and RATIONALE-315 [8].

The primary endpoint extracted was event-free survival (EFS). For each pre-specified clinical subgroup, we extracted the hazard ratio (HR) and its corresponding 95% confidence interval (CI). To satisfy the assumption of normality for model fitting, HRs were natural log-transformed ($ln(\text{HR})$). The standard error (SE) for each $ln(\text{HR})$ was derived from the 95% CI limits based on established meta-analytic methodology [23]:

$$\begin{array}{c}\hfill \text{SE}=\frac{ln(\text{Upper CI})-ln(\text{Lower CI})}{2\times 1.96}\end{array}$$

The variance ($v_i$) was calculated as ${\text{SE}}^2$, and the inverse-variance weight ($w_i$) was defined as $1/v_i$. We also extracted EFS HRs stratified by postoperative pathological responses, specifically pathologic complete response (pCR) and major pathologic response (MPR). For trials where subgroup-specific HRs and 95% CIs were not reported (e.g., the non-pCR cohort in CheckMate 816 and MPR strata in Neotorch), pseudo-individual patient data (IPD) were reconstructed from digitized Kaplan-Meier curves. This was performed using the KM-PoPiGo software (https://kmpopigo.github.io/) to convert survival probabilities into time-to-event data. The reconstructed IPD were then used to estimate the missing HRs and corresponding variances via Cox proportional hazards regression, allowing these cohorts to be included in the meta-regression model. However, to maintain the utility of the machine learning framework as a preoperative stratification tool, these postoperative pathological variables were explicitly excluded from the XGBoost predictive matrix to prevent mediator bias (data leakage).

Multilevel meta-regression framework

We constructed a multilevel meta-regression model using the metafor package [24] in R (version 4.3.3) to synthesize the multidimensional subgroup data and adjust for trial-level baseline imbalances. Because traditional univariable meta-analyses are susceptible to ecological fallacies and the “winner’s curse” in small subgroups, we applied restricted maximum likelihood (REML) estimation with empirical Bayes shrinkage. A nested random-effects structure ($\sim$ 1 | Trial / Subgroup) was specified to account for the hierarchical dependency of subgroup estimates within their respective parent trials. This shrinkage penalizes spurious extreme estimates originating from underpowered subgroups, shrinking them toward the global adjusted mean. This approach provides a conservative estimate of the treatment effect across diverse clinical strata.

Inverse-variance weighted machine learning and SHAP

While multilevel meta-regression harmonizes linear heterogeneity, it requires prior structural assumptions and struggles to capture high-dimensional, non-linear interactions. Applying machine learning to aggregate subgroup data–in the absence of individual patient data (IPD)–conventionally risks severe overfitting and ecological bias. We therefore applied an explainable artificial intelligence (XAI) approach using the eXtreme Gradient Boosting algorithm via the xgboost library (version 1.5.2) in Python (version 3.6.15) [16].

It is critical to note that because the XGBoost model was trained on aggregate summary statistics rather than individual patient data, it strictly captures between-subgroup variance rather than between-patient variance. Consequently, the derived SHAP values estimate the predictive contribution of subgroup characteristics to the mean treatment effect, and should not be directly extrapolated as individual-level biological effects.

The XGBoost regressor was fitted using inverse-variance weights ($w_i$) to mitigate small-sample aggregate bias. This weighting strategy also adjusts for the numerical imbalance between the single neoadjuvant-only trial (CheckMate 816) and the six perioperative trials, ensuring the neoadjuvant cohort is weighted by its statistical precision rather than raw trial count. Overfitting to the limited aggregate data was controlled by restricting the maximum tree depth (max_depth = 3) and setting a low learning rate (learning_rate = 0.05). Categorical features were one-hot encoded prior to model training.

The tree-based ensemble was interpreted using SHapley Additive exPlanations (SHAP) [17] via the shap library (version 0.41.0) in Python to calculate the marginal contribution of each clinical feature to the predicted log HR. SHAP summary plots display the relative importance and directional effect of each clinical feature.

Algorithmic validation and sensitivity analyses

Statistical significance and estimate precision were evaluated through permutation testing and bootstrap resampling. A permutation test ($n=1000$ permutations of the dependent variable, $ln(\text{HR})$) established an empirical null distribution to calculate P-values for feature importance. To assess model stability against data perturbation, we performed bootstrap resampling ($n=1000$ iterations with replacement). This approach allowed us to generate 95% confidence intervals (CIs) for the SHAP estimates and evaluate the robustness of the feature rankings.

Model calibration and algorithmic shrinkage were evaluated using the root mean square error (RMSE), $R^2$, and the calibration slope between predicted and observed $ln(\text{HR})$. A leave-one-trial-out (LOTO) cross-validation assessed the cross-trial robustness and transportability of the feature rankings. To evaluate structural confounding from regional demographics, a sensitivity analysis simultaneously excluded regionally restricted Asian cohorts (e.g., Neotorch and RATIONALE-315).

To evaluate the XGBoost model and assess the necessity of a complex non-linear approach, we compared its performance against two simpler baselines: an Elastic Net and a Random Forest regressor. Both baseline models were fitted using the same inverse-variance weighting strategy ($w_i=1/v_i$) applied to the XGBoost model. Performance was evaluated using the root mean square error (RMSE) and the coefficient of determination ($R^2$).

Synthesis of the hypothesis-generating stratification framework

Top-ranking SHAP features (tumor histology and baseline PD-L1 expression) were structured into an exploratory subgroup stratification framework. The hierarchy of the stratification nodes was determined by the SHAP ordinal ranks. Subgroup-directed treatment hypotheses (e.g., potential suitability for continuous perioperative versus neoadjuvant-only regimens) were derived from the empirical Bayesian effect sizes of the multilevel meta-regression model. The framework was visualized using the DiagrammeR package [25] in R.

Use of large language models

The large language model Kimi (K2) was used solely to improve the grammatical accuracy, clarity, and readability of the manuscript text. The tool was not used to generate any scientific hypotheses, data, results, or references. The authors reviewed, verified, and revised all AI-suggested content and take full responsibility for the integrity and accuracy of the final manuscript.

Results

Study selection and baseline characteristics

A total of seven phase II and III randomized controlled trials (AEGEAN, CheckMate 816, CheckMate 77T, KEYNOTE-671, NADIM II, Neotorch, and RATIONALE-315) evaluating neoadjuvant and perioperative immune checkpoint blockade in resectable non-small cell lung cancer (NSCLC) were included. Baseline trial characteristics and treatment regimens are summarized in Table 1.

Table 1.

Baseline characteristics of the included randomized controlled trials

Trial name	Phase	Setting	Experimental arm (ICI + chemo)	Control arm	Target
AEGEAN [5]	III	Perioperative	Durvalumab + platinum-doublet	Placebo + platinum-doublet	PD-L1
CheckMate 816 [3]	III	Neoadjuvant	Nivolumab + platinum-doublet	Platinum-doublet	PD-1
CheckMate 77T [7]	III	Perioperative	Nivolumab + platinum-doublet	Placebo + platinum-doublet	PD-1
KEYNOTE-671 [4]	III	Perioperative	Pembrolizumab + platinum-doublet	Placebo + platinum-doublet	PD-1
NADIM II [22]	II	Perioperative	Nivolumab + platinum-doublet	Platinum-doublet	PD-1
Neotorch [6]	III	Perioperative	Toripalimab + platinum-doublet	Placebo + platinum-doublet	PD-1

ICI, immune checkpoint inhibitor; chemo, chemotherapy; PD-1, programmed cell death protein 1; PD-L1, programmed death-ligand 1. The treatment setting defines the clinical sequence: Perioperative, indicates immunotherapy administered both pre- and post-surgery; Neoadjuvant, indicates exclusive preoperative administration. All trials use platinum-based doublets as the backbone chemotherapy

Pooled EFS and multilevel meta-regression

Adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) for event-free survival (EFS) across all clinical subpopulations, derived from the multilevel meta-regression model, are presented in Fig. 1.

Fig. 1.

Forest plot of adjusted event-free survival. Adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) for clinical subpopulations were estimated using a multilevel empirical Bayesian model incorporating nested trial-level random effects

EFS benefit was positively associated with PD-L1 expression. High PD-L1 expression ($\ge$50%) corresponded to the greatest EFS improvement (adjusted HR = 0.44, 95% CI 0.33−0.57; $P<0.001$). The intermediate 1–49% stratum showed a moderate benefit (adjusted HR = 0.52, 95% CI 0.41−0.66), and the broad PD-L1 positive cohort ($\ge$1%) had an adjusted HR of 0.47 (95% CI 0.37−0.58). Patients with PD-L1 negative tumors (<1%) exhibited an attenuated benefit (adjusted HR = 0.73, 95% CI 0.59−0.91). No statistically significant EFS improvement was observed in patients with early-stage disease (stage IB/II) (adjusted HR = 0.80, 95% CI 0.44−1.48; $P=0.48$).

Serving as a descriptive prognostic milestone rather than a baseline predictor, EFS benefit varied significantly by postoperative pathological response. Patients achieving a pathological complete response (pCR) showed a substantial EFS improvement (adjusted HR = 0.32, 95% CI 0.12−0.84; $P=0.02$). Although patients without a pCR (non-pCR) maintained a statistically significant benefit over chemotherapy alone, this improvement was attenuated (adjusted HR = 0.72, 95% CI 0.59−0.88; $P=0.001$). A similar pattern was observed for major pathological response (MPR) (adjusted HR = 0.51, 95% CI 0.29−0.89) versus non-MPR (adjusted HR = 0.73, 95% CI 0.59−0.91).

XAI analysis of efficacy variance

Non-linear predictors of event-free survival (EFS) heterogeneity were evaluated using an inverse-variance weighted eXtreme Gradient Boosting (XGBoost) algorithm and interpreted with SHapley Additive exPlanations (SHAP). The SHAP summary plot (Fig. 2) ranks clinical features by importance and shows their directional effects on log-transformed HRs.

Fig. 2.

Predictive signatures of perioperative immunotherapy efficacy. SHAP summary plot based on the inverse-variance weighted XGBoost model. Features are ranked by mean absolute SHAP values. Negative SHAP values denote enhanced EFS benefit (lower HR); positive values denote attenuated benefit

Trial-level variance was identified as the primary driver of efficacy heterogeneity. Specific trials (e.g., Neotorch, AEGEAN, and NADIM II) had the highest mean absolute SHAP values. Permutation testing showed this structural bias was statistically significant ($P<0.05$; Additional file 1: Figure S1). This finding supports the use of a multilevel empirical Bayesian model to account for unmeasured structural effects in the primary meta-regression.

Among subgroup-level clinical covariates, baseline PD-L1 expression was the primary predictive feature. A distinct, non-linear U-shaped predictive pattern was observed (Additional file 1: Figure S2): extreme PD-L1 strata exerted the greatest marginal impact on EFS variance but in opposite directions. Specifically, PD-L1 negativity (<1%) served as a primary driver of attenuated benefit, whereas strong positivity ($\ge$50%) drove enhanced efficacy. Intermediate expression (1–49%) contributed minimally. The predictive significance of these extreme PD-L1 categories, alongside tumor extent (T4), was validated using permutation testing (Additional file 1: Figure S1). While squamous histology and smoking history exhibited substantial mean absolute SHAP values, indicating high clinical impact, their shared variance likely attenuated their independent statistical significance in permutation testing. Nevertheless, their predictive importance was quantified with narrow bootstrap-derived 95% confidence intervals, demonstrating the robustness of these estimates.

Smoking history and histological subtype served as key secondary subgroup-level predictors, outranking demographic and clinical covariates such as sex and nodal status. The "never smoked" subgroup was associated with an attenuated EFS benefit (adjusted HR = 0.67, 95% CI 0.47-−0.95). Subpopulations with squamous cell carcinoma (SCC) histology predicted a greater treatment benefit (adjusted HR = 0.53, 95% CI 0.43−0.65; $P<0.001$), whereas non-squamous cohorts showed positive SHAP values, reflecting an attenuated relative response to perioperative blockade.

A leave-one-trial-out (LOTO) cross-validation assessed the stability of feature importance rankings and cross-regional transportability. This analysis demonstrated that the 1:6 trial imbalance between the neoadjuvant (CheckMate 816) and perioperative settings did not alter the primary feature rankings, indicating their stability across the evidence base, though it cannot fully decouple treatment strategy effects from trial-specific variations. The primary predictive roles of baseline PD-L1 expression, smoking history, and histological subtype remained stable across all individual exclusion scenarios. The prognostic hierarchy was preserved even after the simultaneous exclusion of regionally restricted Asian cohorts (e.g., Neotorch and RATIONALE-315) (Additional file 1: Figure S3).

Comparative benchmarking demonstrated the need for a non-linear approach. The linear Elastic Net underperformed ($R^2=0.162$; RMSE = 0.342), reflecting its inability to capture the U-shaped PD-L1 predictive pattern without explicit feature engineering. While the Random Forest baseline showed improved performance ($R^2=0.248$; RMSE = 0.324), XGBoost achieved the lowest prediction error (RMSE = 0.250) and highest explained variance ($R^2=0.551$) (Table 2). The XGBoost calibration slope of 0.459 indicated a conservative, heavily regularized fit that mitigates small-sample overfitting by systematically compressing predicted outcomes toward the mean. Consequently, model predictions represent exploratory effect-size estimates rather than precise prognostic values.

Table 2.

Performance benchmarking of inverse-variance weighted algorithms

Model architecture	RMSE (Lower is better)	$R^2$ (Higher is better)	Calibration slope
Elastic net (linear baseline)	0.342	0.162	0.134
Random forest (ensemble baseline)	0.324	0.248	0.206
XGBoost (proposed model)	0.250	0.551	0.459

All models were fitted using identical inverse-variance weighting ($w_i=1/v_i$). RMSE, Root Mean Square Error; $R^2$,Coefficient of Determination. Bold values indicate the best performance metrics among the evaluated models

XAI-derived subgroup stratification framework

We synthesized the feature importance rankings and exploratory stratification data into a proposed hypothesis-generating subgroup stratification framework (Fig. 3). This framework integrates routine diagnostic workflows with XAI rankings to stratify the subpopulation by histological subtype and baseline PD-L1 expression, excluding collinear and subjective epidemiological variables (e.g., smoking history) to ensure practical subpopulation stratification. For the squamous cell carcinoma (SCC) cohort, the model indicated greater efficacy with extended therapeutic exposure. Supported by SHAP divergence and exploratory pooled estimates, these findings suggest a continuous perioperative regimen may remain the preferred approach for this subgroup, though cross-strategy comparisons remain confounded by trial-specific variations.

Fig. 3.

XAI-derived hypothesis-generating subgroup stratification framework for perioperative immunotherapy in resectable NSCLC. This exploratory framework integrates XGBoost-SHAP feature importance rankings with empirical stratification data to guide future prospective trial designs. Tumor histology serves as the primary stratification factor: subpopulations with squamous cell carcinoma (SCC) show a statistically significant event-free survival (EFS) benefit from continuous perioperative blockade. For non-squamous cohorts, baseline PD-L1 expression predicts optimal therapeutic duration. Subpopulations with PD-L1 negative (< 1%) tumors require extended perioperative blockade to achieve significant efficacy. PD-L1 positive ($\ge$ 1%) non-squamous tumors demonstrate maximal observed benefit from exclusively neoadjuvant regimens (HR = 0.50 within the CheckMate 816 context), identifying this subgroup as a candidate for adjuvant omission and treatment de-escalation trials. Importantly, the specific hazard ratios presented for combined subgroups (e.g., PD-L1 $\ge$ 1% non-squamous tumors) are model-predicted estimates extrapolated from marginal aggregate data and should be interpreted with caution due to the potential for ecological fallacy. Abbreviations: IO, immunotherapy; SCC, squamous cell carcinoma; HR, hazard ratio; NS, not significant

For non-squamous populations, baseline PD-L1 expression was identified as a potential predictor of optimal therapeutic duration. Subgroups with PD-L1 negative (<1%) non-squamous tumors derived substantial clinical benefit (HR = 0.75) primarily from extended perioperative blockade. XAI signatures for PD-L1 positive ($\ge$1%) non-squamous tumors, however, indicated pronounced responses to neoadjuvant therapy alone. For this cohort, aggregate data models suggested an exclusively neoadjuvant approach yielded the maximal observed benefit (HR = 0.50) within the context of the CheckMate 816 trial.

Within the proposed exploratory framework, adjuvant immunotherapy is conceptually omitted for the PD-L1 positive ($\ge$ 1%) non-squamous subgroup. SHAP analysis demonstrated that prolonging checkpoint blockade into the adjuvant phase yielded attenuated relative efficacy for this subpopulation, identifying them as potential candidates for future treatment de-escalation trials. The resulting framework diverges from uniform adjuvant application, providing a histology- and PD-L1-directed approach to inform subgroup-directed perioperative management and future trial designs.

Discussion

In this synthesis of phase II and III trials, we characterized the sources of heterogeneity in perioperative immunotherapy for NSCLC. While multilevel meta-regression demonstrated a pooled event-free survival (EFS) benefit, the inverse-variance weighted XGBoost-SHAP model revealed that trial-level heterogeneity was the primary driver of variance, outweighing the predictive value of most subgroup-level clinical features. By adjusting for this trial-level confounding, we isolated the key subgroup-level predictors (e.g., baseline PD-L1 expression and histology) to inform subgroup-directed hypotheses for future treatment de-escalation trials.

The identification of specific trials (e.g., Neotorch, AEGEAN, and NADIM II) as primary drivers of EFS heterogeneity underscores a fundamental challenge in evidence synthesis. This trial-level variance likely stems from differing study designs, including variations in chemotherapy backbones, treatment cycles, and placebo use [26]. These findings highlight the limitations of standard aggregate-data meta-analyses, which are susceptible to ecological and aggregation biases [27], thereby supporting our use of a multilevel empirical Bayesian model to adjust for structural imbalances [21]. By treating distinct baseline features and assay methods as selection variables [28], and quantifying trial-specific effects through leave-one-trial-out (LOTO) cross-validation, the XGBoost model mitigates the confounding impact of overarching trial biases on subgroup-level signals. This approach supports the internal validity and external transportability of the identified clinical predictors (e.g., extreme PD-L1 expression and squamous histology) across diverse populations.

Among subgroup-level clinical covariates, baseline PD-L1 expression remained the dominant predictor, albeit with a complex directional influence. Rather than a monotonic biological gradient, we observed a U-shaped importance distribution: extreme PD-L1 strata exerted the greatest marginal impact on model variance but in opposite directions. PD-L1 negativity (<1%) served as a primary driver of attenuated benefit, whereas strong positivity ($\ge$50%) was a primary driver of enhanced efficacy. Smoking history and squamous histology served as key secondary predictors, outranking demographic features such as sex. Biologically, the association between "never smoked" status and attenuated benefit aligns with established mechanisms: tobacco exposure induces a high tumor mutational burden (TMB) characterized by transversion mutations, generating a highly immunogenic, neoantigen-rich tumor microenvironment [29, 30]. Tumors in never-smokers thus often exhibit immune-cold microenvironments, limiting the efficacy of perioperative checkpoint blockade.

These mechanisms align with our finding that squamous cell carcinoma (SCC) histology serves as a primary subgroup-level predictor of enhanced EFS benefit. SCC of the lung is characterized by extensive genomic instability, featuring near-universal TP53 mutations and high somatic mutational burdens linked to heavy smoking [29, 31]. Unlike non-squamous cohorts–which frequently harbor actionable driver alterations (e.g., EGFR or ALK) associated with lower immune infiltration and primary resistance to checkpoint blockade [32, 33]–this pronounced immunogenicity renders SCC subpopulations particularly sensitive to perioperative PD-1/PD-L1 inhibition. As recently noted by Wang et al. [34], the distinct genomic and immune profile of SCC underscores the need for histology-directed subgroup strategies.

While baseline PD-L1 expression and histological subtype are established prognostic features, evaluating them in isolation is insufficient to guide complex sequencing decisions. Rather than merely re-identifying these predictive features, our XAI-derived framework maps their hierarchical relationship, translating them into a hypothesis-generating subgroup stratification framework [10]. To maintain its clinical utility as a preoperative stratification tool, this exploratory framework was restricted to baseline features. While postoperative milestones such as pCR are strong prognostic markers, their integration into initial treatment algorithms is clinically impractical, as these pathological outcomes remain unavailable prior to surgery. It is essential to distinguish these two distinct lines of evidence: our XGBoost-SHAP framework provides baseline-driven predictive associations, whereas pCR/MPR status offers descriptive prognostic stratification but should not be used for causal inference regarding baseline de-escalation. Whereas the data support continuous perioperative blockade for SCC cohorts, the framework identifies a distinct, histology- and PD-L1-directed strategy for non-squamous populations. Specifically, PD-L1 positive ($\ge$1%) non-squamous tumors achieved maximal observed EFS benefit via exclusively neoadjuvant regimens. This proposed omission of the adjuvant phase is supported by cross-tumoral evidence in breast cancer and melanoma, where achieving a pathologic complete response (pCR) diminishes the marginal benefit of further systemic therapy [35, 36]. Our multilevel meta-regression illustrates this prognostic gradient: while achieving pCR is associated with marked EFS improvements, the relative benefit in non-pCR cohorts is substantially attenuated. This survival gap–a challenge similarly described in hepatocellular carcinoma [37]–suggests that the unstratified application of year-long adjuvant immunotherapy may be clinically suboptimal, warranting response-adaptive strategies. Given that our neoadjuvant-specific baseline relies entirely on the CheckMate 816 trial, conclusions regarding definitive treatment de-escalation remain premature. Our findings are strictly hypothesis-generating rather than prescriptive. Although the model identifies PD-L1-positive non-squamous cohorts as potential candidates for adjuvant omission, these predictions should not alter standard clinical practice. Clinical implementation of such response-adaptive strategies requires rigorous validation through prospective, head-to-head adaptive clinical trials [38].

Existing subgroup-level meta-analyses [39, 40] and landmark phase III trials (e.g., IMpower010 and KEYNOTE-091) [41, 42] have evaluated the efficacy of immunotherapy across adjuvant and perioperative settings. While these analyses provide prognostic insights–such as the survival benefit of incorporating preoperative therapy–they typically evaluate clinical features independently or assume linear dose-response relationships. Our study builds on this by capturing complex, non-linear interactions specific to the perioperative and neoadjuvant space. For example, rather than a monotonic PD-L1 dose-response, our model identified a U-shaped predictive pattern: extreme PD-L1 strata ($<1\%$ and $\ge 50\%$) drive the majority of EFS variance, whereas intermediate levels ($1\text{--}49\%$) exert minimal marginal impact. Additionally, by identifying an interaction between tumor histology and PD-L1 expression, our framework shows that PD-L1-positive ($\ge 1\%$) non-squamous cohorts experience an attenuated relative benefit from the adjuvant phase, highlighting them as potential candidates for treatment de-escalation trials [10].

Our approach integrates machine learning with epidemiological evidence synthesis [43]. Conventional aggregate-data meta-analyses in highly heterogeneous fields like perioperative oncology are susceptible to ecological fallacies [27, 44]. To address cross-trial baseline disparities, we incorporated empirical Bayes shrinkage [21] and applied inverse-variance weighting to our XGBoost model using subgroup-level estimates. This framework overcomes the limitations of traditional univariable random-effects models by directly accounting for trial-level confounding, thus avoiding the statistical power loss associated with a priori stratification [45].

This study has several limitations. First, our dataset features a 1:6 trial-level imbalance between the neoadjuvant-only (CheckMate 816) and perioperative cohorts, reflecting the current availability of phase III evidence. To mitigate potential bias during model training, we applied inverse-variance weighting. Subsequent LOTO sensitivity analyses showed that this imbalance did not skew the primary feature rankings. However, since our neoadjuvant-specific baseline relies entirely on the CheckMate 816 trial, our model cannot fully decouple true neoadjuvant-specific effects from CheckMate-816 trial-specific variations (such as distinct scheduling, drug-specific factors, or patient demographics). Similarly, a direct comparison between PD-1 and PD-L1 inhibitors was not feasible. Given that AEGEAN was the only included trial evaluating a PD-L1 inhibitor, including drug target as a feature would cause perfect collinearity with the trial indicator. Although we absorbed these drug-specific variations into the trial-level random effects, prospective head-to-head trials are required to assess true differential efficacy between targets. Second, the lack of an independent external validation cohort precludes a definitive assessment of model generalizability. As our analysis synthesized all currently available mature phase II and III trials, no independent dataset is currently available for this purpose. While our LOTO framework provided internal-external cross-validation, future validation with independent prospective cohorts remains necessary. Third, relying on trial-level summary statistics inherently aggregates heterogeneous patient outcomes, potentially obscuring nuanced treatment responses within complex clinical subgroups [46]. Furthermore, synthesizing aggregate data leaves the analysis vulnerable to ecological bias and Simpson’s paradox [47]. The specific hazard ratios presented in our subgroup stratification framework (Fig. 3) for combined subgroups (e.g., PD-L1 $\ge$ 1% non-squamous tumors) are model-predicted estimates extrapolated from marginal aggregate data, and should be interpreted with caution. Additionally, the XGBoost model yielded a calibration slope of 0.46. While this reflects a conservative, regularized fit that penalizes extreme observed values to prevent overfitting, it also indicates a systematic compression of predicted outcomes toward the mean. Consequently, the model’s predicted values must be interpreted as exploratory effect-size estimates rather than precise prognostic scores for individual clinical decision-making. Future validation using large-scale individual patient data (IPD)–derived from real-world cohorts or reconstructed from published Kaplan-Meier curves [48]–is required to confirm these findings and inform individualized clinical decisions.

Conclusions

In summary, our inverse-variance weighted machine learning framework demonstrates that a non-linear, U-shaped PD-L1 predictive pattern and squamous histological subtype are robust subgroup-level predictors of immunotherapy efficacy in resectable NSCLC. While permutation testing validates the independent predictive significance of extreme PD-L1 strata and primary tumor extent (T4), SHAP sensitivity analyses reveal that histology and smoking history maintain highly stable predictive rankings across diverse clinical trials. Clinically, given the reliance on aggregate data and trial-level confounding, these integrated findings provide a hypothesis-generating framework rather than prescriptive guidelines. For patients with PD-L1-positive ($\ge$1%) non-squamous tumors, our exploratory analyses suggest that these individuals may be potential candidates for future adaptive trials evaluating neoadjuvant-only approaches, underscoring the need for prospective research to validate these stratification hypotheses and safely omit adjuvant treatment to minimize unnecessary toxicity.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AD

Aggregate-data

ALK

Anaplastic lymphoma kinase

Chemo

Chemotherapy

CI

Confidence interval

EFS

Event-free survival

EGFR

Epidermal growth factor receptor

HR

Hazard ratio

ICI

Immune checkpoint inhibitor

IO

Immunotherapy

IPD

Individual patient data

irAEs

Immune-related adverse events

LOTO

Leave-one-trial-out

ML

Machine learning

NS

Not significant

NSCLC

Non-small cell lung cancer

pCR

Pathological complete response

PD-1

Programmed cell death protein 1

PD-L1

Programmed death-ligand 1

REML

Restricted maximum likelihood

RMSE

Root mean square error

SCC

Squamous cell carcinoma

SE

Standard error

SHAP

SHapley Additive exPlanations

TMB

Tumor mutational burden

XAI

Explainable artificial intelligence

XGBoost

EXtreme Gradient Boosting

Author contributions

FL and ZZ conceived the study and designed the overall analytical framework. FS and JL conducted the systematic literature search, performed data extraction, and curated the clinical subgroup datasets. ZZ developed the methodology, implemented the machine learning algorithms (inverse-variance weighted XGBoost and SHAP), executed the multilevel meta-regression, and generated the data visualizations. FS and JL provided clinical interpretation of the U-shaped predictive patterns and assisted in structuring the exploratory subgroup stratification framework. FS and ZZ drafted the original manuscript. FL supervised the project, provided clinical oversight, and critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript.

Funding

This work was supported by the Shanxi Key Laboratory of Tumor Organoids (grant no. 202504010931054). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Data availability

All data analyzed during this study are included in this published article and its supplementary information files. The aggregate-level summary statistics (hazard ratios and 95% confidence intervals) utilized in our analytical framework were systematically extracted from the primary publications and conference reports of the following trials: AEGEAN [5], CheckMate 816 [3], CheckMate 77T [7], KEYNOTE-671 [4], NADIM II [22], Neotorch [6], and RATIONALE-315 [8]. The compiled input dataset is available within the supplementary files and the associated GitHub repository. The analytical codebase, encompassing the R scripts for multilevel meta-regression (metafor v4.3.3) and the Python scripts (v3.6.15) for the inverse-variance weighted XGBoost (v1.5.2) and SHAP (v0.41.0) framework, is publicly accessible to ensure full computational reproducibility. The complete code and datasets are available at https://github.com/ZheqingZhu/Meta-SHAP.

Declarations

Conflict of interest

The authors declare that they have no conflict of interest.

Consent for publication

Not applicable.

Ethical approval

Not applicable. This study was based on the synthesis of publicly available, aggregate-level statistics from previously published and ethically approved clinical trials. No individual patient data were directly accessed or utilized; therefore, additional ethics committee approval and patient consent were not required.