The Association and Predictive Value of Nutritional and Inflammatory Biomarkers in Advanced Non-Small Cell Lung Cancer Response to Immune Checkpoint Inhibitors

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dekker et al. analyzed blood nutritional and inflammatory biomarkers before initiation of immune checkpoint inhibitors in NSCLC patients. A total of 555 patients were included in this study. The researchers found that elevated levels of C-reactive protein (CRP) and high Glasgow Outcome Scale (GPS) scores were associated with poor treatment response or even with patient’s death. On the other hand, higher levels of Prognostic Nutritional Index (PNI) and Advanced Lung Cancer Inflammatory Index (ALI) indicated better survival and tumor control. The study demonstrated that by measuring a few simple blood tests, doctors can identify patients at higher risk and make better treatment decisions to provide more individualized care to patients with NSCLC.

 

Breakthroughs in the study of the prediction of immunotherapy in cases of advanced NSCLC by Dekker et al.

1. Innovative Definition of Response Assessment (Modified RECIST).

Breakthrough: This study also included patients with “disease stability (SD) lasting 6 months or longer” in the responder population.

Innovative Value: The research team believes this approach more accurately reflects clinically relevant disease control, particularly in assessing short-term primary endpoints.

2. Focus on Early Prognosis and Mortality Risk

Almost all of the existing studies focus on the long-term data such as OS and PFS.

Breakthrough: This study specifically focuses on “mortality within three months” and early treatment response.

Innovative Value: Early detection of mortality risk or primary drug resistance in individual patients enables their doctors to change their treatment as soon as necessary and to provide the best possible individualized care in the long run.

3. The use of a large number of single inflammatory markers was not the focus of this study. Rather, the investigators validated and applied a number of integrated physiological states such as ALI and PNI.

Breakthrough: Particular emphasis was placed on the predictive power of ALI and PNI.

Innovative Value: PNI = a measure of a patients' nutrition status and their immune efficacy (measured by lymphocyte count), which makes it a particularly valuable marker in the context of ICI treatment.

ALI, which combines information on BMI, nutrition and systemic inflammation, appears to be a more comprehensive indicator for the prediction of early disease progression.

4.Rigorous Statistical Validation Methods

Breakthrough: The authors employed a multivariate logistic regression model with “bootstrapping” for internal validation of our findings.

Innovative Value: Through 1000 iterations of analysis the stability of the markers to predict early progression in cancer patients was confirmed. In a multivariate model, including age, sex and clinical stage, the blood tests CRP, PNI and ALI proved to be strong early predictors for clinical outcome.

In summary, this study defines clinically effective treatment response and sets up a predictive model for early risk management of immunotherapy in advanced lung cancer using well-established and inexpensive blood markers.

Although the current document has a lot of clinical value, there are some issues that need to be solved.

Limitations of the study design:

1. The non-standard definition of response, which the researchers have coined “Modified ORR”. This would include patients with stable disease for 6 months or longer in addition to the classical cases of CR and partial PR. Note that the Modified ORR is not in line with the standard RECIST definition of ORR, where only CR and PR are counted as responders.

A serious point of contention is the non-standard definition of response to treatment, using the term ‘responder’ to describe patients with Modified ORR (as defined in the article). In clinical practice, the RECIST criteria are used to determine the change in tumor size during treatment. According to these criteria, patients are classified as CR or PR, whereas patients with stable disease are not considered to have responded to treatment. The use of the term ‘responder’ in this study, however, includes patients with stable disease for as long as 6 months. This definition of ‘responder’ is at odds with standard clinical practice and may not allow for comparison with the results of other studies.

2. Sample Representativeness and Generalizability:

Point of contention: The major points of criticism for this study are the selection of a patient group with a poor prognosis and that results are based on a single center (UMCG Hospital in the Netherlands), a tertiary referral center. Although researchers have noted that these factors might have overestimated the power of these predictors in early treatment response, they limit clinical applicability and do not ensure that results are generalizable to other patient groups with different stages of disease or to other healthcare systems.

3. Lack of External Validation: Using bootstrapping of 1000 iterations for the internal validation of the study’s findings is not enough for establishing the clinical reliability of the results.

Comment: This breakthrough study needs to be further validated on independent external patient cohorts to confirm the clinical reliability of this and other models.

4. Lack of Time-to-Event Data: The study contains single time points of measurement (i.e. three months after the start of treatment). The model cannot be used for a dynamic prognostic prediction and it is not possible to say whether the indicators at baseline quickly lose their predictive power or not. A time-to-event analysis (e.g. by means of Kaplan-Meier curves) would be more appropriate.

Question: What is the value of the model for follow-up of patients over time using time-to- event data, for example Kaplan-Meier curves, for assessment of progression-free survival?

5.Statistical Collinearity Issues

When building multivariate models to predict outcomes of patients, often variables are left out of the final model because of high collinearity with other included variables. In this study albumin was left out of the final model for the multivariate prediction, because of high collinearity with the composite variables PNI, GPS and ALI. It is unclear what the independent value of pure albumin would be for the prediction of outcome in this patient group.

Concerns: This could make it difficult to clarify the independent effect of pure albumin on the prediction of patient outcome, or the effect of certain other prognostic indicators could be obscured by the results of the composite scores.

 

The study described is another example of a simple, cost-effective clinical tool for cancer prognosis. As for any study, however, there are some limits to the results that can be expected and the way these results are presented. This study is a retrospective study, which implies that there is no external validation yet for the results found. In addition, some of the definitions of the variables used for the models are not standard, which might make comparison with results from other studies using different definitions difficult.

 

More concerns:

1. Unexplained symbol insertion: In multiple definitions and formulas sentences within definitions and formulas are enclosed in incorrect **「」∗∗ symbols. Examples include the end of the sentence in “disease ≥ 6 months”, “after the start of treatment…$”, and numerous other locations throughout the document. The same problem exists in the repeated abbreviation descriptions found at the end of the document.
2. The words three-moth instead of three-month are misspelled throughout the paper (e.g. 111 and 313).
3. Inconsistencies in the notation of percentage values: While in the text percentages are written with a comma in some places (e.g. “69,7%”, “83,8%” in line 111 and 313), in tables (e.g. Table 3) the same values are written with a period in the same places (e.g. “83.4%”).
4. Omitted Spaces: For example, in “Figure S1: Receiver” above, a space is missing after the colon.
5. Terminology Misspelling: In the list of abbreviations for terms such as Receiver operating characteristic is misspelled as Receiving operating characteristic.

6.The Ambiguity of the Definition of "Significance":

The study has defined in the abstract, the following parameters as significant predictors for progression in the ORR analysis: CRP, GPS, PNI, and ALI. Surprisingly, in the bootstrapping analysis for the internal validation of the ORR model in Table 5, only ALI (p-value: 0.03) is a statistically significant predictor (p < 0.05) while CRP (p-value: 0.10), GPS (p-value: 0.09), and PNI (p-value: 0.09) are not.

7.Logical overlap in endpoints:

As mentioned previously, all patients in this study had advanced cancer (i.e. stage III or IV). It is well known that the predictive ability of various clinical and biological indicators could be overestimated in such a patient group and that the results obtained may not be generalizable to earlier-stage cancer patients. In addition, the current study did not validate the results obtained in an independent external cohort of patients. Therefore, the clinical reliability of the model obtained in this study is questionable.

8.Inconsistent descriptions of adjusted variables:

Lines 234-235 state that age, sex, comorbidities, clinical stage and monotherapy were not included in the mortality model since they were not significant. However, lines 239-240 state that the ORR model was adjusted for clinical stage and monotherapy. This appears to be a confusion on the part of the author, since the two models are addressing different endpoints. Without careful reading of the table notes, the reader is left wondering which variables were included in the model.

Author Response

Response to Reviewer 1 Comments   Dear reviewer,   We would like to thank you for the careful evaluation of our manuscript. We think you provided constructive comments and suggestions. In response to the comments, we have thoroughly revised the manuscript and addressed all points. We believe that these revisions have substantially strengthened the paper by improving the research design, methods, results presentation and interpretation of the findings. In addition, we reviewed the English language throughout the manuscript to improve readability. The results changed slightly compared with the previous version because we developed a revised model during the revision process. Detailed responses to each comment are provided below, and all corresponding changes are shown using track changes in the revised manuscript.   We hope that the revisions address the concerns raised and hope the manuscript will be suitable for publication.   Sincerely,   Erick Suazo-Zepeda, MD, PhD Graduate School of Medical Sciences, University of Groningen Zusterhuis, Hanzeplein 1, Groningen, The Netherlands Tel: +31 50 361 0738/9 E-Mail: e.suazo.zepeda@umcg.nl

Point by point response   Comment 1: The non-standard definition of response, which the researchers have coined “Modified ORR”. This would include patients with stable disease for 6 months or longer in addition to the classical cases of CR and partial PR. Note that the Modified ORR is not in line with the standard RECIST definition of ORR, where only CR and PR are counted as responders. A serious point of contention is the non-standard definition of response to treatment, using the term ‘responder’ to describe patients with Modified ORR (as defined in the article). In clinical practice, the RECIST criteria are used to determine the change in tumor size during treatment. According to these criteria, patients are classified as CR or PR, whereas patients with stable disease are not considered to have responded to treatment. The use of the term ‘responder’ in this study, however, includes patients with stable disease for as long as 6 months. This definition of ‘responder’ is at odds with standard clinical practice and may not allow for comparison with the results of other studies

Response 1: We agree with the reviewer that our use of the term modified ORR, which included stable disease for at least 6 months, differs from the standard RECIST definition, where only CR and PR are considered as objective responses. We acknowledge that this definition may limit the comparability to other studies. To address this concern, we performed a sensitivity analysis in which we used the standard RECIST definition of ORR, which confirmed that our main findings were not affected by this distinction. Moreover, we changed the terminology from modified ORR to disease control rate (DCR), to avoid ambiguity.   Change in text: Objective response rate (ORR) is changed to disease control rate (DCR) throughout the complete manuscript.   Discussion: Moreover, we used a DCR that included patients with stable disease lasting at least six months, as well as those achieving PR or CR. This differs from the conventional definition of objective response rate (ORR), which includes only PR and CR. This approach was chosen because durable disease stabilization was considered a clinically meaningful outcome in our study population, given the median survival of patients with metastatic NSCLC ranges from 3 – 34 months [26]. Importantly, a sensitivity analysis using the standard definition of ORR was also performed and yielded consistent results, supporting the robustness of our findings.   Methods: 2.4.4 Sensitivity analysis A sensitivity analysis was performed using a standard RECIST-based 3-month progression endpoint. The same multivariable logistic regression model used in the primary analysis was fitted, including Glasgow Prognostic Score, Advanced Lung Cancer Inflammation Index, and Prognostic Nutritional Index, adjusted for age, sex, clinical stage, treatment regimen, treatment line, and comorbidities. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC).   Results: 3.3.5. Sensitivity analysis Using the RECIST-based endpoint, only 4 of 505 patients (0.8%) were reclassified compared with the primary analysis. Results were comparable to the primary model. Higher Glasgow Prognostic Score was associated with increased odds of 3-month progression (OR 1.67, 95% CI 1.05–2.66), whereas higher Advanced Lung Cancer Inflammation Index (OR 0.99, 95% CI 0.97–1.00) and Prognostic Nutritional Index (OR 0.93, 95% CI 0.87–0.99) were associated with lower odds of 3-month progression (Table S4). Model discrimination was similar to the primary analysis (AUC 0.732 vs. 0.729) (Figure S3).

Comment 2: Sample Representativeness and Generalizability: Point of contention: The major points of criticism for this study are the selection of a patient group with a poor prognosis and that results are based on a single center (UMCG Hospital in the Netherlands), a tertiary referral center. Although researchers have noted that these factors might have overestimated the power of these predictors in early treatment response, they limit clinical applicability and do not ensure that results are generalizable to other patient groups with different stages of disease or to other healthcare systems.

Response 2: We agree with the reviewer that the inclusion of patients treated at a tertiary referral center may limit the generalizability of our findings, particularly for those with more favorable prognoses. However, during the early implementation of ICI treatment in 2015-2018, treatment was centralized in the UMCG. As a result, the majority of patients in our cohort from this period likely represents the real-world population receiving ICIs at that time. Patients treated later (2019 – 2021) may reflect a more selected and potentially more complex subgroup, as treatment was provided by other centers as well. We have explained this in the discussion.   Changes in text: Discussion: As the UMCG is a tertiary referral hospital, its patient population predominantly consists of individuals with advanced disease. Consequently, our cohort may not be fully representative of patients treated in secondary care settings. This may have led to an overestimation of the predictive performance of the nutritional and inflammatory markers, thereby limiting the generalizability of our findings to patients with a more favorable prognosis. However, during the early implementation phase of ICIs (2015-2018), the treatment was centralized at the UMCG. As a result, patients treated during this period likely reflect a broader real-world population, whereas those treated between 2019-2021 may represent a more selected and clinically complex subgroup.    Comment 3: Lack of External Validation: Using bootstrapping of 1000 iterations for the internal validation of the study’s findings is not enough for establishing the clinical reliability of the results. Comment: This breakthrough study needs to be further validated on independent external patient cohorts to confirm the clinical reliability of this and other models.   Response 3: We agree that internal validation through bootstrapping does not replace external validation. Internal bootstrapping is only useful for evaluating the performance of our model within this specific cohort, and it cannot establish the generalizability of the model to other patient populations or clinical settings.   Accordingly, we revised the discussion and conclusion sections to emphasize that external validation in independent cohorts is required before the model can be considered for routine clinical use. Nevertheless, the biomarkers included in our model have previously been associated with treatment outcomes in patients with advanced NSCLC, supporting their biological and clinical relevance. We emphasized that external validation in independent patient cohorts is necessary before our models can be considered for routine clinical implementation.   Changes in text: Discussion: Although internal validation indicated that the models performed consistently within our population, external validation is necessary to determine their generalizability to independent patient cohorts.   Conclusion: Baseline biomarkers routinely measured in standard blood draws, including PNI and ALI, were associated with three-month mortality. In the final multivariable model, lower PNI and lower ALI remained independently predictive of mortality in our cohort of patients with NSCLC treated with ICIs. Similarly, higher CRP, higher GPS, lower PNI and lower ALI were predictive of disease progression.   However, these findings should be externally validated in independent patient cohorts before implementation in routine clinical practice can be considered.    Comment 4: Lack of Time-to-Event Data: The study contains single time points of measurement (i.e. three months after the start of treatment). The model cannot be used for a dynamic prognostic prediction and it is not possible to say whether the indicators at baseline quickly lose their predictive power or not. A time-to-event analysis (e.g. by means of Kaplan-Meier curves) would be more appropriate. Question: What is the value of the model for follow-up of patients over time using time-to- event data, for example Kaplan-Meier curves, for assessment of progression-free survival?   Response 4: We thank the reviewer for this comment. We agree that the use of a fixed endpoint of three-months does not provide information on how the predictive value of the biomarkers evolves over time. However, the primary objective of the present study was to identify patients at risk of early mortality and treatment response within three months after initiation of ICI treatment. This timeframe was chosen because a substantial number of patients with advanced NSCLC experience early mortality despite treatment with ICIs. If we are able to identify those patients before start of treatment, we may avoid burden from treatment with limited expected benefit during the final stage of life. Time-to-event analyses would have allowed for evaluation of the timing of progression or death over the entire follow-up period, which includes whether the biomarkers remain consistently prognostic over time. We have added a paragraph to the discussion that addresses our rationale for the three-moth endpoint and the limitations of not using time-to-event analysis.   Changes in text: Discussion: This study focused on short-term outcomes, namely three-month mortality and treatment response after initiation of ICI treatment. By using a fixed three-month endpoint, patients experiencing progression or death after this time point were classified as non-events, which may underestimate associations with longer-term outcomes. Therefore, the findings should be interpreted as associations with short-term outcomes rather than long-term prognosis.   Comment 5: Statistical Collinearity Issues When building multivariate models to predict outcomes of patients, often variables are left out of the final model because of high collinearity with other included variables. In this study albumin was left out of the final model for the multivariate prediction, because of high collinearity with the composite variables PNI, GPS and ALI. It is unclear what the independent value of pure albumin would be for the prediction of outcome in this patient group. Concerns: This could make it difficult to clarify the independent effect of pure albumin on the prediction of patient outcome, or the effect of certain other prognostic indicators could be obscured by the results of the composite scores. Response 5: We thank the reviewer for this comment. We agree that several biomarkers considered in this study are mathematically related. In particular, albumin is incorporated in both PNI and ALI, while inflammatory components overlap across several biomarkers. Consequently, multicollinearity was carefully re-evaluated during model development. In response to this comment, we assessed VIFs and correlations between candidate biomarkers. Based on these analyses and a more stringent evaluation of collinearity, CRP was removed from the final prediction model because it showed substantial overlap with the information captured by the composite indices. Albumin was also not included as an independent predictor because it is a component of both PNI and ALI. The final models therefore retained only biomarkers that contributed independently to model performance while demonstrating acceptable collinearity diagnostics. The VIF values and biomarker correlation matrix have been added to the Supplementary Materials, and the Methods and Results sections have been updated accordingly. Changes in text: Methods: Multicollinearity was assessed using the generalized variance inflation factor (GVIF), with values >5 considered indicative of problematic collinearity. Complete-case analyses were performed.   Results: 3.3.3 Multicollinearity No evidence of problematic multicollinearity was observed in either final model. Adjusted generalized variance inflation factor (GVIF^(1/(2×Df))) values ranged from 1.03 to 1.32 in the final 3-month progression model and from 1.03 to 1.11 in the final 3-month mortality model, indicating minimal correlation among predictors and stable coefficient estimates (Table S3).   Comment 6:  1.Unexplained symbol insertion: In multiple definitions and formulas sentences within definitions and formulas are enclosed in incorrect **「」∗∗ symbols. Examples include the end of the sentence in “disease ≥ 6 months”, “after the start of treatment…$”, and numerous other locations throughout the document. The same problem exists in the repeated abbreviation descriptions found at the end of the document. 2.The words three-moth instead of three-month are misspelled throughout the paper (e.g. 111 and 313). 3.Inconsistencies in the notation of percentage values: While in the text percentages are written with a comma in some places (e.g. “69,7%”, “83,8%” in line 111 and 313), in tables (e.g. Table 3) the same values are written with a period in the same places (e.g. “83.4%”). 4.         Omitted Spaces: For example, in “Figure S1: Receiver” above, a space is missing after the colon. 5.         Terminology Misspelling: In the list of abbreviations for terms such as Receiver operating characteristic is misspelled as Receiving operating characteristic. Response 6: We have corrected inconsistencies and corrected spelling and formatting errors throughout the manuscript.   Comment 7:  The Ambiguity of the Definition of "Significance": The study has defined in the abstract, the following parameters as significant predictors for progression in the ORR analysis: CRP, GPS, PNI, and ALI. Surprisingly, in the bootstrapping analysis for the internal validation of the ORR model in Table 5, only ALI (p-value: 0.03) is a statistically significant predictor (p < 0.05) while CRP (p-value: 0.10), GPS (p-value: 0.09), and PNI (p-value: 0.09) are not.   Response 7: We thank the reviewer for this comment. We agree that the inclusion of p-values in the original version of Table 5 may have led to confusion regarding the purpose of the bootstrap analysis. In the initial manuscript, p-values were recorded alongside the bootstrap results. However, the aim of this analysis was not to reassess the statistical significance of the predictors, but to evaluate the internal validity and stability of the model estimates through bootstrap resampling. We used bootstrap validation to assess the degree of optimism and reliability of the estimated effects across repeated samples. The statistical significance of the biomarkers was already evaluated in the final multivariable model. Therefore, to avoid ambiguity, we removed the p-values from Table 5.   Change in text: Methods: 2.4.3 Bootstrap Internal validation was conducted using 1,000 bootstrap resamples. Optimism-corrected AUCs were calculated by subtracting the average optimism from the apparent AUC. Bootstrap distributions were used to evaluate the stability of regression coefficients and derive bootstrap-based 95% confidence intervals.   Results: 3.3.4 Bootstrap Internal validation using 1,000 bootstrap samples demonstrated good stability of the final models. For the 3-month progression model, the apparent AUC was 0.735 and the optimism-corrected AUC was 0.711, indicating limited overfitting. The bootstrap estimates confirmed the associations observed in the original model, with higher GPS (OR 1.79, 95% CI 1.10–2.99), lower ALI (OR 0.99, 95% CI 0.97–1.00), and lower PNI (OR 0.92, 95% CI 0.86–0.99) remaining associated with increased odds of 3-month progression (Table 5). For the 3-month mortality model, the apparent AUC was 0.817 and the optimism-corrected AUC was 0.791. The bootstrap analysis confirmed the stability of the biomarker effects, with lower ALI (OR 0.96, 95% CI 0.94–0.99) and lower PNI (OR 0.83, 95% CI 0.76–0.89) remaining independently associated with increased odds of 3-month mortality (Table 5). No substantial changes in the direction or magnitude of the regression coefficients were observed across bootstrap samples.   Comment 8:  Logical overlap in endpoints: As mentioned previously, all patients in this study had advanced cancer (i.e. stage III or IV). It is well known that the predictive ability of various clinical and biological indicators could be overestimated in such a patient group and that the results obtained may not be generalizable to earlier-stage cancer patients. In addition, the current study did not validate the results obtained in an independent external cohort of patients. Therefore, the clinical reliability of the model obtained in this study is questionable.   Response 8: We refer to our response to comments two and three. We agree that the inclusion of only patients with advanced-stage NSCLC may limit the applicability of our findings to patients with earlier-stage disease. However, the present study was specifically designed to develop prediction models for patients with advanced NSCLC receiving immune checkpoint inhibitors, which represents the clinical population for whom these models are intended. We further acknowledge that the lack of external validation limits the assessment of the model’s generalizability and clinical reliability. Therefore, external validation in independent cohorts is required before these models can be considered for routine clinical application.   Comment 9:  Inconsistent descriptions of adjusted variables: Lines 234-235 state that age, sex, comorbidities, clinical stage and monotherapy were not included in the mortality model since they were not significant. However, lines 239-240 state that the ORR model was adjusted for clinical stage and monotherapy. This appears to be a confusion on the part of the author, since the two models are addressing different endpoints. Without careful reading of the table notes, the reader is left wondering which variables were included in the model.   Response 9: We have corrected the inconsistent descriptions.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript by Dekker et al. represents a methodologically sound retrospective cohort study addressing a clinically relevant question: whether routine blood-based inflammatory and nutritional biomarkers can predict early outcomes (3-month mortality and objective response rate) in NSCLC patients receiving ICIs. The sample size (n=505) is substantial, and the use of bootstrapping for internal validation is a strength. However, several issues should be addressed before publication.

- The discrimination (AUC 0.81 for mortality, 0.73 for progression) is moderate at best. Clinicians would likely need a higher AUC or a validated risk threshold to change management. The authors should discuss what threshold of probability would justify, for example, withholding ICI or switching therapy.

- Three-month mortality as an endpoint is appropriate for early prediction, but the authors should clarify whether this is all-cause mortality (likely yes, given the definition on page 2). Competing risks (e.g., treatment toxicity vs. cancer progression) are not discussed.

- The definition of ORR is non-standard. The authors classified stable disease (SD) ≥ 6 months as "response." This deviates substantially from RECIST criteria, where ORR includes only complete response (CR) and partial response (PR). This should be relabeled as "disease control rate (DCR)" or "clinical benefit rate" to avoid confusion. Using the term ORR is misleading and limits comparability with other studies.

- PD-L1 expression is missing from all multivariable models. PD-L1 is a well-established predictor of ICI response. The authors collected this variable (page 3) but did not include it in the final models. Even if non-significant or with many missing values, this should be reported and justified. Its exclusion is a major limitation.

- Collinearity among biomarkers is insufficiently addressed. Albumin was excluded due to high collinearity, but PNI contains albumin and ALI contains both albumin and NLR. The final model includes CRP, PNI, and ALI together. The authors should report variance inflation factors (VIFs) or a correlation matrix to demonstrate that collinearity does not distort the estimates.

- Handling of missing data is not described. Missingness is present for several biomarkers (e.g., CRP missing in 4.5% of alive patients, 1.2% of deceased). The authors do not state whether they used complete-case analysis, multiple imputation, or another method. This should be specified.

- No external validation is performed despite the claim to "validate" in the introduction (page 2). Bootstrapping provides internal validation only. The authors should rephrase claims of validation and acknowledge that external validation in an independent cohort is needed.

- The confidence intervals for GPS in the mortality model are extremely wide (OR 24.54, 95% CI: 9.74–61.83). These estimates are unstable. The authors should interpret them with caution and consider whether sample sizes in GPS subgroups are sufficient for reliable estimation.

- The OR for CRP in the ORR model is 1.00 (1.00–1.01). This is statistically borderline and clinically negligible. The text should tone down claims of "significance" for this association.

- The abstract, particularly the results section, is overloaded with abbreviations (e.g., OR, CI, CRP, PNI, ALI, GPS, ORR), making it difficult to follow. I suggest spelling out terms at first use or avoiding abbreviations altogether in the abstract.

- The main results are presented in tabular form. I believe that histograms would be more informative for key findings, while the original tables could be moved to the supplementary materials.

Author Response

Response to Reviewer 2 Comments

 

Dear reviewer,

 

We would like to thank you for the careful evaluation of our manuscript. We think you provided constructive comments and suggestions. In response to the comments, we have thoroughly revised the manuscript and addressed all points. We believe that these revisions have substantially strengthened the paper by improving the research design, methods, results presentation and interpretation of the findings. In addition, we reviewed the English language throughout the manuscript to improve readability.

The results changed slightly compared with the previous version because we developed a revised model during the revision process.  

Detailed responses to each comment are provided below, and all corresponding changes are shown using track changes in the revised manuscript.

 

We hope that the revisions address the concerns raised and hope the manuscript will be suitable for publication.

 

Sincerely,

 

Erick Suazo-Zepeda, MD, PhD

Graduate School of Medical Sciences, University of Groningen

Zusterhuis, Hanzeplein 1, Groningen, The Netherlands

Tel: +31 50 361 0738/9

E-Mail: e.suazo.zepeda@umcg.nl

 

Point by point response

 

Comment 1: The discrimination (AUC 0.81 for mortality, 0.73 for progression) is moderate at best. Clinicians would likely need a higher AUC or a validated risk threshold to change management. The authors should discuss what threshold of probability would justify, for example, withholding ICI or switching therapy.

 

Response 1: We thank the reviewer for this important comment. We agree that an AUC of 0.82 and 0.73 are not sufficient to support direct clinical decision-making. In this study, no risk-threshold was predefined for guiding treatment decisions such as withholding or switching ICI therapy, since the primary aim was to develop and internally validate a prognostic model rather than to define clinical decision rules.

We acknowledge that clinical implementation would require externally validated risk thresholds and more research to determine clinically relevant cut-offs. We have added this point to the discussion.

 

Change in text:

Discussion: Despite showing moderate discrimination (AUC of 0.82 for mortality and 0.74 for progression), the model is not yet suitable as a clinical decision tool, since no externally validated risk thresholds were defined to guide treatment decisions.

 

 

 

 

Comment 2: Three-month mortality as an endpoint is appropriate for early prediction, but the authors should clarify whether this is all-cause mortality (likely yes, given the definition on page 2). Competing risks (e.g., treatment toxicity vs. cancer progression) are not discussed.

 

Response 2: Three-month mortality was defined as all-cause mortality occurring within three months after initiation of ICI treatment, and this has now been clarified in the Methods section.

 

We agree that competing risks, such as death related to treatment toxicity, cancer progression, or other causes, may influence mortality outcomes. However, cause-specific mortality data were not available in our dataset, and the primary objective of this study was to predict short-term mortality irrespective of the underlying cause of death. Therefore, competing risks were not analyzed separately. We have added this limitation to the Discussion section.

 

Change in text:

Methods: Three-month mortality was defined as death by any cause three months after treatment initiation.

 

Discussion: Moreover, the lack of cause-specific mortality data implies that different causes of death, including disease progression, treatment-related toxicity and comorbidities could not be distinguished. Treatment-related toxicity could also indirectly reflect treatment response and could therefore contain clinically relevant information about prognosis.

 

Comment 3: The definition of ORR is non-standard. The authors classified stable disease (SD) ≥ 6 months as "response." This deviates substantially from RECIST criteria, where ORR includes only complete response (CR) and partial response (PR). This should be relabeled as "disease control rate (DCR)" or "clinical benefit rate" to avoid confusion. Using the term ORR is misleading and limits comparability with other studies.

 

Response 3: We agree with the reviewer and adjusted the terminology throughout the manuscript. We changed ORR to DCR to avoid ambiguity. Moreover,  we performed a sensitivity analysis in which we used the standard RECIST definition of ORR, which confirmed that our main findings were not affected by this distinction, which showed that there is no difference in results. This additional analysis makes sure that this research is comparable to other studies.

 

Changed in text:
Objective response rate (ORR) is changed to disease control rate (DCR) throughout the complete manuscript.

 

Discussion: Moreover, we used a DCR that included patients with stable disease lasting at least six months, as well as those achieving PR or CR. This differs from the conventional definition of objective response rate (ORR), which includes only PR and CR. This approach was chosen because durable disease stabilization was considered a clinically meaningful outcome in our study population, given the median survival of patients with metastatic NSCLC ranges from 3 – 34 months [26]. Importantly, a sensitivity analysis using the standard definition of ORR was also performed and yielded consistent results, supporting the robustness of our findings.

 

Methods: 2.4.4 Sensitivity analysis

A sensitivity analysis was performed using a standard RECIST-based 3-month progression endpoint. The same multivariable logistic regression model used in the primary analysis was fitted, including Glasgow Prognostic Score, Advanced Lung Cancer Inflammation Index, and Prognostic Nutritional Index, adjusted for age, sex, clinical stage, treatment regimen, treatment line, and comorbidities. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC).

 

Results: 3.3.5. Sensitivity analysis

Using the RECIST-based endpoint, only 4 of 505 patients (0.8%) were reclassified compared with the primary analysis. Results were comparable to the primary model. Higher Glasgow Prognostic Score was associated with increased odds of 3-month progression (OR 1.67, 95% CI 1.05–2.66), whereas higher Advanced Lung Cancer Inflammation Index (OR 0.99, 95% CI 0.97–1.00) and Prognostic Nutritional Index (OR 0.93, 95% CI 0.87–0.99) were associated with lower odds of 3-month progression (Table S4). Model discrimination was similar to the primary analysis (AUC 0.732 vs. 0.729) (Figure S3).

 

Comment 4: PD-L1 expression is missing from all multivariable models. PD-L1 is a well-established predictor of ICI response. The authors collected this variable (page 3) but did not include it in the final models. Even if non-significant or with many missing values, this should be reported and justified. Its exclusion is a major limitation.

 

Response 4: We thank the reviewer for this important comment. We agree that PD-L1 expression is an established predictor of response to immune checkpoint inhibitors and should be considered when developing predictive models.

Although PD-L1 expression was collected, it was unavailable in 161 of 505 patients (31.9%) because PD-L1 testing was not routinely performed during a considerable part of the study period (2015–2017), particularly in the earlier years of ICI implementation. As a result, inclusion of PD-L1 in the multivariable models would have reduced the effective sample size by approximately one-third or required the inclusion of a large “unknown” category, limiting the interpretability and usefulness of the variable within the prediction model.

We have clarified this in the manuscript and acknowledge the inability to adequately evaluate the contribution of PD-L1 as a limitation of the study. Future studies with more complete PD-L1 data should assess its additional predictive value.

 

Change in text:

Discussion: Due to the study period, PD-L1 expression was unavailable in 161 of 505 patients (31.9%), as PD-L1 testing was not routinely performed. Currently, PD-L1 is the most important biomarker in ICI treatment, and exclusion of this biomarker in our study could have influenced the completeness of available predictive information.

 

Comment 5: Collinearity among biomarkers is insufficiently addressed. Albumin was excluded due to high collinearity, but PNI contains albumin and ALI contains both albumin and NLR. The final model includes CRP, PNI, and ALI together. The authors should report variance inflation factors (VIFs) or a correlation matrix to demonstrate that collinearity does not distort the estimates.

Response 5:

We thank the reviewer for this comment. We agree that several biomarkers considered in this study are mathematically related. In particular, albumin is incorporated in both PNI and ALI, while inflammatory components overlap across several biomarkers. Consequently, multicollinearity was carefully re-evaluated during model development.

In response to this comment, we assessed VIFs and correlations between candidate biomarkers. Based on these analyses and a more stringent evaluation of collinearity, CRP was removed from the final prediction model because it showed substantial overlap with the information captured by the composite indices. Albumin was also not included as an independent predictor because it is a component of both PNI and ALI.

The final models therefore retained only biomarkers that contributed independently to model performance while demonstrating acceptable collinearity diagnostics. The VIF values and biomarker correlation matrix have been added to the Supplementary Materials, and the Methods and Results sections have been updated accordingly.

Change in text:
Methods: Multicollinearity was assessed using the generalized variance inflation factor (GVIF), with values >5 considered indicative of problematic collinearity. Complete-case analyses were performed.

Results: 3.3.3 Multicollinearity

No evidence of problematic multicollinearity was observed in either final model. Adjusted generalized variance inflation factor (GVIF^(1/(2×Df))) values ranged from 1.03 to 1.32 in the final 3-month progression model and from 1.03 to 1.11 in the final 3-month mortality model, indicating minimal correlation among predictors and stable coefficient estimates (Table S3).

Comment 6: Handling of missing data is not described. Missingness is present for several biomarkers (e.g., CRP missing in 4.5% of alive patients, 1.2% of deceased). The authors do not state whether they used complete-case analysis, multiple imputation, or another method. This should be specified.

 

Response 6: We thank the reviewer or this comment. The method section has been further clarified accordingly. We have now explained that we conducted complete-case analysis.

 

Change in text:

Methods: Complete-case analyses were performed. 

 

Comment 7: No external validation is performed despite the claim to "validate" in the introduction (page 2). Bootstrapping provides internal validation only. The authors should rephrase claims of validation and acknowledge that external validation in an independent cohort is needed.

 

 

 

Response 7: We thank the reviewer for this thoughtful comment. We have revised the introduction and replaced the term validate with internally validate to accurately reflect the study design. We have also clarified in the discussion that external validation in an independent cohort is required to validate the model.

 

Change in text:

Introduction: Therefore, to contribute to the decision-making process regarding ICI treatment for lung cancer, this study aims to evaluate and internally validate the association between routinely measured biomarkers and composite indices with both three-month mortality and treatment response assessed using a modified disease control rate (DCR) definition based on RECIST categories, in which stable disease lasting ≥6 months was classified as disease control.   

 

Discussion: Although internal validation indicated that the models performed consistently within our population, external validation is necessary to determine their generalizability to independent patient cohorts.

 

Conclusion:

Baseline biomarkers routinely measured in standard blood draws, including lower PNI, and lower ALI were associated with and jointly predictive of three-month mortality in our cohort of patients with NSCLC treated with ICIs. Similarly, higher CRP, higher GPS, lower PNI and lower ALI were predictive of disease progression.

 

However, these findings should be externally validated in independent patient cohorts before implementation in routine clinical practice can be considered. 

 

Comment 8: The confidence intervals for GPS in the mortality model are extremely wide (OR 24.54, 95% CI: 9.74–61.83). These estimates are unstable. The authors should interpret them with caution and consider whether sample sizes in GPS subgroups are sufficient for reliable estimation.

 

Response 8: We agree with the reviewer that the confidence intervals for GPS in the mortality model are wide. This probably reflects the limited number of events within the GPS 2 group. We have revised the Discussion to emphasize the uncertainty around this estimate and to caution against overinterpretation. We also clarify that the limited number of events in the subgroup likely contributed to the instability.

 

Change in text:

Discussion: However, the estimate for GPS 2 should be interpreted with caution, as the small subgroup size resulted in greater uncertainty, which is reflected by the wide CI.

 

Comment 9: The OR for CRP in the ORR model is 1.00 (1.00–1.01). This is statistically borderline and clinically negligible. The text should tone down claims of "significance" for this association.

 

Response 9: We thank the reviewer for this comment. Following the revision of the biomarker analyses and model development strategy, the results for CRP have changed and the association is no longer characterized by an odds ratio close to unity with borderline statistical significance. In the revised analyses, higher CRP was associated with both 3-month progression (OR 1.01, 95% CI 1.01–1.02; p<0.001) and 3-month mortality (OR 1.02, 95% CI 1.01–1.02; p<0.001) in univariable analyses.

However, CRP was not retained in the final multivariable prediction models because of overlap with the composite biomarkers and concerns regarding collinearity. The manuscript has been revised accordingly, and the interpretation of CRP has been updated to reflect the revised analyses.

 

Within the new multivariable regression model, ALI shows a CI of 0.97 – 1.00. The actual higher CI is 0.998, so this number is shown as 1 but is in fact smaller.

 

Comment 10: The abstract, particularly the results section, is overloaded with abbreviations (e.g., OR, CI, CRP, PNI, ALI, GPS, ORR), making it difficult to follow. I suggest spelling out terms at first use or avoiding abbreviations altogether in the abstract.

Response 10: We thank the reviewer for this comment. We respectfully disagree that the abstract is overloaded with abbreviations. All abbreviations are defined at first use in accordance with journal guidelines, and several of the terms (e.g., OR and CI) are commonly used in biomedical abstracts. Nevertheless, the abstract has been revised following the updated biomarker selection and final models, and we have reviewed the text to ensure clarity and readability.

Change in text:
Results: Among 505 patients, 421 were alive and 84 deceased; 297 were responders and 208 non-responders. In the final progression model, higher GPS was associated with increased odds of 3-month progression, whereas higher ALI (OR 0.99, 95% CI 0.97–1.00) and higher PNI (OR 0.93, 95% CI 0.87–0.99) were associated with decreased odds of 3-month progression. Higher ALI (OR 0.97, 95% CI 0.94–0.99), and higher PNI (OR 0.83, 95% CI 0.78–0.89) were associated with lower odds of 3-month mortality. The mortality-model showed an AUC of 0.82 and 0.74 for the disease progression model. Sensitivity analysis with the standard RECIST definition revealed similar results.   

Comment 11: The main results are presented in tabular form. I believe that histograms would be more informative for key findings, while the original tables could be moved to the supplementary materials.

Response 11: We thank the reviewer for this suggestion. We agree that graphical displays can be useful for illustrating data distributions and key findings. However, we believe that the current tables provide the most complete presentation of the multivariable model results, including odds ratios, confidence intervals, and p-values, which are essential for interpretation of the study findings.

For this reason, we have chosen to keep the tables in the main manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

• Cohort is from 2015–2020 — up to 6 years old. ICI regimens, combination strategies, and patient selection have changed substantially since then. Authors must address how this limits generalisability.
• Paper claims clinical relevance but offers no actionable thresholds or cut-offs. If the audience is clinicians, what should they do when CRP is high or PNI is low? Provide explicit guidance or revise the framing.
• N=505 from a single tertiary centre is modest for 8 candidate predictors. The deceased subgroup (N=84) is small — overfitting risk remains even with bootstrapping. No power calculation or sample size justification is provided.
• Six comorbidities (diabetes, hypertension, COPD, rheumatological disease, dementia, CVD) collapsed into one binary variable. COPD and rheumatological conditions independently alter inflammatory markers and ICI outcomes — this aggregation needs justification or individual modelling.
• Discussion opens with mechanistic speculation before restating what was actually found. Key results (AUC 0.81, specific ORs) should come first, interpretation second.
• GPS for ORR (OR=1.52, CI 0.94–2.45, p=0.09) and PNI for ORR (OR=0.94, CI 0.88–1.01, p=0.09) both cross the null — yet both are described as "significant predictors" in the abstract and conclusions. This must be corrected.
• Rationale for three-month mortality as the primary endpoint is weak. Why not six-month or one-year? The choice needs empirical or clinical justification.
• Albumin dropped for collinearity but VIF values are not reported and the clinical implications of its exclusion are not discussed — albumin is a core marker in this patient group.
• No external validation. Internal bootstrapping is insufficient for a single-centre, dated cohort. Claims of clinical utility should be tempered until an independent cohort confirms the findings.
• AUC values (0.81 mortality, 0.73 ORR) — the most clinician-relevant numbers — are absent from the abstract.
• Up to 9.5% missing data for ALI. Method for handling missingness is not stated.
• Including SD ≥6 months as "response" deviates from standard RECIST ORR. A sensitivity analysis using the standard definition is needed.
• Baseline biomarker values are not stratified by line of treatment in Table 1. First-line vs. later-line ICI patients likely differ meaningfully in both profiles and outcomes.

Author Response

Response to Reviewer 3 Comments

 

Dear reviewer,

 

We would like to thank you for the careful evaluation of our manuscript. We think you provided constructive comments and suggestions. In response to the comments, we have thoroughly revised the manuscript and addressed all points. We believe that these revisions have substantially strengthened the paper by improving the research design, methods, results presentation and interpretation of the findings. In addition, we reviewed the English language throughout the manuscript to improve readability.

The results changed slightly compared with the previous version because we developed a revised model during the revision process.

Detailed responses to each comment are provided below, and all corresponding changes are shown using track changes in the revised manuscript.

 

We hope that the revisions address the concerns raised and hope the manuscript will be suitable for publication.

 

Sincerely,

 

Erick Suazo-Zepeda, MD, PhD

Graduate School of Medical Sciences, University of Groningen

Zusterhuis, Hanzeplein 1, Groningen, The Netherlands

Tel: +31 50 361 0738/9

E-Mail: e.suazo.zepeda@umcg.nl

 

Point by point response

 

Comment 1: Cohort is from 2015–2020 — up to 6 years old. ICI regimens, combination strategies, and patient selection have changed substantially since then. Authors must address how this limits generalizability.

 

Response 1: We thank the reviewer for this important comment. The cohort dates from 2015 – 2020, a period in which ICI treatment strategies have evolved considerable. We have now added a statement in the discussion to explicitly acknowledge that these developments could limit the generalizability of our findings to the current clinical practice.

 

Change in text:

Discussion: Also, ICI strategies have evolved substantially, including increased use of combination regimens and changes in patient selection. Therefore, while our findings support the prognostic value of baseline inflammatory and nutritional biomarkers, their direct generalizability to current ICI treatment regimens should be interpreted with caution.

 

Comment 2: Paper claims clinical relevance but offers no actionable thresholds or cut-offs. If the audience is clinicians, what should they do when CRP is high or PNI is low? Provide explicit guidance or revise the framing.

 

 

Response 2: We thank the reviewer for this insightful comment. We have revised the manuscript to avoid implication of this model in the clinics. Instead, we now explicitly state that the models are not intended to provide clinical decision thresholds at this stage. We further emphasize that external validation in independent cohorts is essential before these biomarkers can be translated into routine clinical decision-making.

 

Change in text:

Discussion: Despite showing moderate discrimination (AUC of 0.82 for mortality and 0.73 for progression), the model is not yet suitable as a clinical decision tool, since no externally validated risk thresholds were defined to guide treatment decisions.

 

Conclusion: These findings suggest that inflammatory and nutritional biomarkers may contribute to risk stratification in patients treated with ICIs. However, these findings should be externally validated in independent patient cohorts before implementation in routine clinical practice can be considered.    

 

Comment 3: N=505 from a single tertiary centre is modest for 8 candidate predictors. The deceased subgroup (N=84) is small — overfitting risk remains even with bootstrapping. No power calculation or sample size justification is provided.

 

Response 3: We thank the reviewer for this comment. We agree that the relatively limited number of events, particularly for 3-month mortality, increases the risk of overfitting when developing prediction models. To minimize this risk, we restricted the number of predictors included in the final models through a structured model selection process based on model fit, discrimination, and collinearity diagnostics. In addition, internal validation was performed using 1,000 bootstrap resamples to evaluate model optimism and coefficient stability.

 

As this was a retrospective study using an existing cohort, no formal sample size calculation was performed prior to the analyses. We acknowledge that the sample size and number of events may have limited the complexity of the models that could be reliably developed and have added this as a limitation in the Discussion section. External validation in larger independent cohorts remains necessary to further assess model performance and generalizability.

 

Change in text:

Discussion: Another limitation of the study is its retrospective study design. The available sample size and number of events may have limited the complexity of the models that could be reliably developed. Although internal validation suggested that the models performed consistently within our population, external validation is required to determine whether our findings can be generalized to independent patient cohorts.

 

Comment 4: Six comorbidities (diabetes, hypertension, COPD, rheumatological disease, dementia, CVD) collapsed into one binary variable. COPD and rheumatological conditions independently alter inflammatory markers and ICI outcomes — this aggregation needs justification or individual modelling.

Response 4: We thank the reviewer for this comment. We agree that specific comorbidities, particularly COPD and rheumatological disease, may independently influence inflammatory biomarkers and treatment outcomes. However, the number of progression and mortality events in our cohort limited the number of predictors that could be included in the multivariable models without increasing the risk of overfitting. To preserve model stability and comply with recommended events-per-variable considerations, comorbidities were therefore combined into a single binary variable indicating the presence of one or more comorbid conditions.

We acknowledge that this approach may have obscured differences in the effects of individual comorbidities and have added this as a limitation in the Discussion section. Future studies with larger sample sizes and a greater number of events should evaluate the independent contribution of specific comorbid conditions.

Change in text:
Discussion: Also, due to limited availability of detailed data on comorbidities, individual conditions such as COPD and rheumatological disease could not be modelled separately and were therefore combined in one single binary comorbidity, which may have masked condition specific effects on inflammatory markers and outcomes. 

Comment 5: Discussion opens with mechanistic speculation before restating what was actually found. Key results (AUC 0.81, specific ORs) should come first, interpretation second.

 

Response 5: We agree with the reviewer and adjusted our discussion accordingly.

 

Change in text:

Discussion: Both models demonstrated good discriminatory performance, with an AUC of 0.82 for mortality and 0.74 for disease progression. Sensitivity analysis showed consistent results with the primary model.  

 

Comment 6: GPS for ORR (OR=1.52, CI 0.94–2.45, p=0.09) and PNI for ORR (OR=0.94, CI 0.88–1.01, p=0.09) both cross the null — yet both are described as "significant predictors" in the abstract and conclusions. This must be corrected.

 

Response 6: We thank the reviewer for this comment. We agree that the presentation of the bootstrap analysis in the original manuscript may have caused confusion regarding the interpretation of statistical significance.

The purpose of the bootstrap analysis was not to reassess statistical significance, but to evaluate internal validity, stability and optimism of the model estimates through repeated resampling. In the revised manuscript, we therefore removed p-values from Table 5 and now present the estimates with a 95% CI only.

Moreover, we reperformed analyses with a different biomarker composition. Therefore, the results changed.

 

 

Change in text:

Methods: 2.4.3 Bootstrap

Internal validation was conducted using 1,000 bootstrap resamples. Optimism-corrected AUCs were calculated by subtracting the average optimism from the apparent AUC. Bootstrap distributions were used to evaluate the stability of regression coefficients and derive bootstrap-based 95% confidence intervals.

 

Results: 3.3.4 Bootstrap

Internal validation using 1,000 bootstrap samples demonstrated good stability of the final models. For the 3-month progression model, the apparent AUC was 0.735 and the optimism-corrected AUC was 0.711, indicating limited overfitting. The bootstrap estimates confirmed the associations observed in the original model, with higher GPS (OR 1.79, 95% CI 1.10–2.99), lower ALI (OR 0.99, 95% CI 0.97–1.00), and lower PNI (OR 0.92, 95% CI 0.86–0.99) remaining associated with increased odds of 3-month progression (Table 5).

For the 3-month mortality model, the apparent AUC was 0.817 and the optimism-corrected AUC was 0.791. The bootstrap analysis confirmed the stability of the biomarker effects, with lower ALI (OR 0.96, 95% CI 0.94–0.99) and lower PNI (OR 0.83, 95% CI 0.76–0.89) remaining independently associated with increased odds of 3-month mortality (Table 5). No substantial changes in the direction or magnitude of the regression coefficients were observed across bootstrap samples.

 

Comment 7: Rationale for three-month mortality as the primary endpoint is weak. Why not six-month or one-year? The choice needs empirical or clinical justification.

 

Response 7: We thank the reviewer for this comment.

The primary objective of the present study was to identify patients at risk of early mortality and treatment response within three months after initiation of ICI treatment. This timeframe was chosen because a substantial number of patients with advanced NSCLC experience early mortality despite treatment with ICIs. If we are able to identify those patients before start of treatment, we may avoid burden from treatment with limited expected benefit during the final stage of life.

We have added a paragraph to the discussion that addresses our rationale for the three-moth endpoint and the limitations.

 

Change in text:
This study focused on short-term outcomes, namely three-month mortality and treatment response after initiation of ICI treatment. By using a fixed three-month endpoint, patients experiencing progression or death after this time point were classified as non-events, which may underestimate associations with longer-term outcomes. Therefore, the findings should be interpreted as associations with short-term outcomes rather than long-term prognosis. Moreover, we used a DCR that included patients with stable disease lasting at least six months, PR or CR. This differs from the standard definition of objective response rate (ORR), which only includes PR and CR. This approach was chosen due to durable disease stabilization was considered a clinically meaningful outcome in our study population, with the median survival of patients metastasized NSCLC ranging from 3 – 34 months [26].

 

 

 

Comment 8: Albumin dropped for collinearity but VIF values are not reported and the clinical implications of its exclusion are not discussed — albumin is a core marker in this patient group.

Response 8: We thank the reviewer for this comment. We agree that several biomarkers considered in this study are mathematically related. In particular, albumin is incorporated in both PNI and ALI, while inflammatory components overlap across several biomarkers. Consequently, multicollinearity was carefully re-evaluated during model development.

In response to this comment, we assessed VIFs and correlations between candidate biomarkers. Based on these analyses and a more stringent evaluation of collinearity, CRP was removed from the final prediction model because it showed substantial overlap with the information captured by the composite indices. Albumin was also not included as an independent predictor because it is a component of both PNI and ALI.

The final models therefore retained only biomarkers that contributed independently to model performance while demonstrating acceptable collinearity diagnostics. The VIF values and biomarker correlation matrix have been added to the Supplementary Materials, and the Methods and Results sections have been updated accordingly.

Changes in text:

Methods: Multicollinearity was assessed using the generalized variance inflation factor (GVIF), with values >5 considered indicative of problematic collinearity. Complete-case analyses were performed.

 

Results: 3.3.3 Multicollinearity

No evidence of problematic multicollinearity was observed in either final model. Adjusted generalized variance inflation factor (GVIF^(1/(2×Df))) values ranged from 1.03 to 1.32 in the final 3-month progression model and from 1.03 to 1.11 in the final 3-month mortality model, indicating minimal correlation among predictors and stable coefficient estimates (Table S3).

 

Comment 9: No external validation. Internal bootstrapping is insufficient for a single-centre, dated cohort. Claims of clinical utility should be tempered until an independent cohort confirms the findings.

 

Response 9: We want to thank the reviewer again for this thoughtful comment. We also agree that internal validation through bootstrapping does not replace external validation. Internal bootstrapping is only useful for evaluating the performance of our model within this specific cohort. Therefore, we revised the discussion and conclusion. We emphasized that external validation in independent patient cohorts is necessary before our models can be considered for routine clinical implementation.

 

Change in text

Discussion: Although internal validation indicated that the models performed consistently within our population, external validation is necessary to determine their generalizability to independent patient cohorts.

 

 

Conclusion: Baseline biomarkers routinely measured in standard blood draws, including lower PNI, and lower ALI were associated with and jointly predictive of three-month mortality in our cohort of patients with NSCLC treated with ICIs. Similarly, higher GPS, lower PNI and lower ALI were predictive of disease progression

 

However, these findings should be externally validated in independent patient cohorts before implementation in routine clinical practice can be considered. 

 

Comment 10: AUC values (0.81 mortality, 0.73 ORR) — the most clinician-relevant numbers — are absent from the abstract.

 

Response 10: We thank the reviewer for this comment and agree that this information needs to be added to the abstract. We did add them. The AUCs changed because of the new models.

 

Change in text:

Abstract: The mortality-model showed an AUC of 0.82 and 0.73 for the disease progression model. Sensitivity analysis with the standard RECIST definition revealed similar results.

 

Comment 11:  Up to 9.5% missing data for ALI. Method for handling missingness is not stated.

 

Response 11: We thank the reviewer or this comment. The method section has been further clarified accordingly, by adding that complete case analyses were performed.  

 

Change in text: Complete-case analyses were performed.

 

Comment 12: Including SD ≥6 months as "response" deviates from standard RECIST ORR. A sensitivity analysis using the standard definition is needed.

 

Response 12:  We agree with the reviewer that our definition deviates from standard RECIST ORR, therefore performed a sensitivity analysis in which we used the standard RECIST definition of ORR, which confirmed that our main findings were not affected by this distinction. Moreover, we changed the terminology from modified ORR to disease control rate (DCR), to avoid ambiguity. Please also see our answer to comment 3 of reviewer 1.

 

Change in text

Objective response rate (ORR) is changed to disease control rate (DCR) throughout the complete manuscript.

 

Discussion: Moreover, we used a DCR that included patients with stable disease lasting at least six months, PR or CR. This differs from the standard definition of objective response rate (ORR), which only includes PR and CR. This approach was chosen due to durable disease stabilization was considered a clinically meaningful outcome in our study population, with the median survival of patients metastasized NSCLC ranging from 3 – 34 months [26]. Importantly, sensitivity analysis which used the standard definition of ORR was performed, which supports the generalizability of our findings.

 

Methods: 2.4.4 Sensitivity analysis

A sensitivity analysis was performed using a standard RECIST-based 3-month progression endpoint. The same multivariable logistic regression model used in the primary analysis was fitted, including Glasgow Prognostic Score, Advanced Lung Cancer Inflammation Index, and Prognostic Nutritional Index, adjusted for age, sex, clinical stage, treatment regimen, treatment line, and comorbidities. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC).

 

Results: 3.3.5. Sensitivity analysis

Using the RECIST-based endpoint, only 4 of 505 patients (0.8%) were reclassified compared with the primary analysis. Results were comparable to the primary model. Higher Glasgow Prognostic Score was associated with increased odds of 3-month progression (OR 1.67, 95% CI 1.05–2.66), whereas higher Advanced Lung Cancer Inflammation Index (OR 0.99, 95% CI 0.97–1.00) and Prognostic Nutritional Index (OR 0.93, 95% CI 0.87–0.99) were associated with lower odds of 3-month progression (Table S4). Model discrimination was similar to the primary analysis (AUC 0.732 vs. 0.729) (Figure S3).

 

Comment 13: Baseline biomarker values are not stratified by line of treatment in Table 1. First-line vs. later-line ICI patients likely differ meaningfully in both profiles and outcomes.

Response 13: We thank the reviewer for this comment. To address this concern, we performed an additional analysis comparing baseline biomarker values across treatment lines and have included the results as a Supplementary Table S2. We observed statistically significant differences in several baseline biomarkers, including CRP, albumin, PNI, and GPS, across treatment lines. In particular, patients receiving later-line treatment tended to have less favorable inflammatory and nutritional profiles than those receiving first-line treatment.

To account for these differences, treatment line was included as an adjustment variable in all multivariable prediction models. Furthermore, the supplementary analysis has been added to provide a more detailed description of baseline biomarker distributions according to treatment line.

Change in text
Methods: In addition, baseline biomarker values between treatment lines were assessed with Kruskal-Wallis test for continuous variables and chi-square test for categorical variables.

Results: Several baseline biomarkers differed significantly between treatment lines, including CRP, albumin, GPS and PNI. CRP values were highest in patient receiving third-line treatment. PNI levels were similar in treatment lines 2 and 3, 40.0 (38.0–43.0) vs 40.0 (36.0–42.0), whereas slightly higher values were observed in first-line treatment: 41.0 (38.0–44.0). Between all treatment lines, GPS 1 was the most frequently observed category (50.6% vs 57.8% vs 59.2%) (Table S2).

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

After carefully review,the questions were answered adequately and  I really appreciated authors' comprehensive reply. I have no more questions

Reviewer 2 Report

Comments and Suggestions for Authors

I appreciate the authors for their revisions. The manuscript has been improved and, in my opinion, is now suitable for publication.

Reviewer 3 Report

Comments and Suggestions for Authors

 

No more comments