Glucose-6-Phosphate Dehydrogenase Deficiency and Cardiovascular Risk in Familial Hypercholesterolemia: A Retrospective Cohort Study

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Brief Summary

This retrospective cohort study examines the association between glucose-6-phosphate dehydrogenase (G6PD) deficiency and cardiovascular disease risk in 217 patients with familial hypercholesterolemia from Sardinia. The authors report that G6PD-deficient FH patients had significantly higher CVD prevalence. The study's main strength lies in addressing a novel gene-gene interaction in a unique population with high prevalence of both conditions, providing the first evidence that G6PD deficiency may act as a cardiovascular risk modifier in FH.

Comments Regarding General Concepts

The manuscript addresses an important clinical question with a reasonable study design for exploratory research. However, several methodological limitations compromise the strength of conclusions. The retrospective design cannot establish temporal relationships between G6PD deficiency and CVD events, which is critical for causal inference. Most importantly, the mechanistic hypothesis linking G6PD deficiency to increased oxidative stress lacks direct biochemical validation through measurement of relevant biomarkers.

The study's reliance on clinical FH diagnosis for subset of patients without genetic confirmation introduces potential misclassification bias. While the authors acknowledge this limitation, it could significantly impact the validity of gene-gene interaction claims. 

Specific Comments

Lines 29: The claim that "baseline LDL-C did not differ by G6PD status" contradicts results in Table 2 showing significantly higher LDL-C in G6PD-deficient patients (356.0 vs 282.9 mg/dL, p=0.002).

Table 2: The age difference between groups (58.2 vs 44.5 years, p=0.004) represents a unneglectable factor that may explain much of the CVD difference, as CVD prevalence increases dramatically with age.

Recommendations

1. Provide direct measurement of oxidative stress biomarkers to validate the proposed mechanism
2. Conduct age-stratified analyses or propensity score matching to address the significant age imbalance
3. Acknowledge limitations more prominently and moderate clinical recommendations accordingly

The manuscript presents intriguing preliminary evidence for an important clinical association but requires additional validation and mechanistic data before definitive conclusions can be drawn about clinical management implications.

Author Response

This retrospective cohort study examines the association between glucose-6-phosphate dehydrogenase (G6PD) deficiency and cardiovascular disease risk in 217 patients with familial hypercholesterolemia from Sardinia. The authors report that G6PD-deficient FH patients had significantly higher CVD prevalence. The study's main strength lies in addressing a novel gene-gene interaction in a unique population with high prevalence of both conditions, providing the first evidence that G6PD deficiency may act as a cardiovascular risk modifier in FH.

We thank the reviewer for the constructive suggestions. Below, we address each comment in turn. Our replies in blue follow the reviewers’ comments. All changes in the revised manuscript are highlighted as requested.

Comments 1: The manuscript addresses an important clinical question with a reasonable study design for exploratory research. However, several methodological limitations compromise the strength of conclusions. The retrospective design cannot establish temporal relationships between G6PD deficiency and CVD events, which is critical for causal inference. Most importantly, the mechanistic hypothesis linking G6PD deficiency to increased oxidative stress lacks direct biochemical validation through measurement of relevant biomarkers

Response 1: We acknowledge the reviewer’s concern and agree that causal inferences cannot be drawn from a retrospective study design. This limitation has been explicitly noted in the revised version of the manuscript (page 12, lines 426-427). We also concur that our study did not directly assess oxidative stress biomarkers to demonstrate the mechanistic link between G6PD deficiency and increased oxidative stress (page 12, lines 445–447). Nevertheless, this hypothesis is strongly supported by a substantial body of prior literature (e.g., PMID: 39142558; 35578850; 35402109; 34949182; 34138756; 33502933; 33050491; 30731288; 29626298; 28823591), which consistently reports that G6PD deficiency is associated with increased oxidative stress.

Comment 2: The study's reliance on clinical FH diagnosis for subset of patients without genetic confirmation introduces potential misclassification bias. While the authors acknowledge this limitation, it could significantly impact the validity of gene-gene interaction claims.

Response 2: We appreciate the reviewer’s insightful comment. We fully agree that genetic testing represents the gold standard for diagnosing FH, but acknowledge that it is not universally performed in clinical practice. In our cohort, patients without available genetic results were included if they fulfilled established clinical diagnostic criteria (e.g., Simon Broome). In the revised manuscript, we now explicitly clarify in both the Methods and Discussion sections that a subset of FH diagnoses was based on clinical rather than genetic criteria (page 3, lines 114-115). As highlighted in the literature, approximately 30–40% of clinically diagnosed FH patients do not carry an identifiable pathogenic mutation (PMID: 26844336; PMID: 36422206). Consequently, we cannot exclude some degree of misclassification (e.g., polygenic hypercholesterolemia being labeled as FH). While clinical criteria are validated and widely used, we recognize that the absence of genetic confirmation may introduce heterogeneity. This limitation is now clearly acknowledged in the revised Discussion and Limitations sections (page 12, lines 431–435).

Comment 3: Lines 29: The claim that "baseline LDL-C did not differ by G6PD status" contradicts results in Table 2 showing significantly higher LDL-C in G6PD-deficient patients (356.0 vs 282.9 mg/dL, p=0.002).

Response 3: We thank the reviewer for pointing out this inconsistency. Upon re-examining our data, we identified an error in the manuscript text. As correctly reported in Table 2, LDL-C levels were significantly higher in the G6PD-deficient group compared with the non-deficient group, and this difference reached statistical significance in the univariate analysis. We have corrected the abstract to accurately state this finding (mean ± SD: 356 ± 96 mg/dL vs. 283 ± 77 mg/dL, p = 0.002; page 1, lines 27–29; page 8, lines 264-266). We apologize for the oversight and have carefully reviewed the manuscript to ensure consistency between the text and the reported results. No changes were required for other biomarker values.

Comment 4: Table 2: The age difference between groups (58.2 vs 44.5 years, p=0.004) represents a unneglectable factor that may explain much of the CVD difference, as CVD prevalence increases dramatically with age.

Response 4: We thank the reviewer for underscoring the importance of age as a potential confounder. Indeed, G6PD-deficient patients in our cohort were, on average, older than non-deficient patients. To address this imbalance, our primary logistic regression model included age as a covariate, thereby adjusting for its effect. In addition, as recommended, we performed supplementary analyses to further exclude the possibility that age differences accounted for the observed association. Specifically, we conducted propensity score–matched analyses (1:1, 1:2, and 1:4 matching on age, sex, BMI, smoking status, and hypertension, using nearest-neighbor matching without replacement). After matching, the groups had comparable age distributions, yet G6PD deficiency remained significantly associated with higher CVD prevalence. These additional results are now detailed in the Supplementary Material. We have further noted in the Discussion that the observed association persisted after rigorous adjustment for age, consistent with prior findings (e.g., Pes et al., PMID: 30731288). We believe these additions strengthen the manuscript by demonstrating that the G6PD–CVD association is not merely an artifact of age imbalance.

Comment 5: Provide direct measurement of oxidative stress biomarkers to validate the proposed mechanism.

Response 5: We agree with the reviewer that direct measurement of oxidative stress biomarkers would substantially strengthen mechanistic claims. Our study did not include such measurements (e.g., glutathione, malondialdehyde [MDA], or related markers), and therefore we cannot provide new data in this regard. This limitation is now explicitly acknowledged in the revised Discussion (page 12, lines 445–447). While we hypothesize that oxidative stress contributes to the increased CVD risk in G6PD-deficient patients, we recognize that this remains unproven in our cohort. To provide context, we cite recent studies reporting biochemical evidence of elevated oxidative stress in G6PD deficiency, such as Andrews et al. (PMID: 40292229), who demonstrated increased lipid and protein oxidation products (MDA and protein carbonyls) in affected individuals. Accordingly, we have tempered our language throughout the manuscript, replacing statements implying established oxidative stress elevation with more cautious phrasing (e.g., “may have” or “hypothesized to have”), and explicitly noting the absence of direct biomarker data in our study. These revisions make our conclusions more balanced and aligned with the available evidence.

Comment 6: Conduct age-stratified analyses or propensity score matching to address the significant age imbalance

Response 6: We appreciate the reviewer’s suggestion to conduct age-stratified analyses and propensity score matching to account for the observed age imbalance. While the relatively small size of the G6PD-deficient subgroup (n = 31) initially discouraged us from applying propensity score methods, we performed these analyses in response to the reviewer’s comment.

	PS* matched 1:1		PS matched 1:2		PS matched 1:4
	G6PD# deficient	G6PD normal	G6PD deficient	G6PD normal	G6PD deficient	G6PD normal
No. subjects	31	31	31	62	31	124
Male, %	9 (29.0)	9 (29.0)	9 (29.0)	17 (27.4)	9 (29.0)	45 (36.3)
Mean age, y (SD)	57.9 (15.8)	57.4 (13.5)	57.9 (15.8)	57.7 (13.9)	57.9 (15.8)	55.7 (13.7)
Mean BMI, kg/m² (SD)	26.4 (3.9)	24.6 (3.2)	26.4 (3.9)	25.1 (3.3)	26.4 (3.9)	24.8 (3.4)
Current or former smokers, %	14 (45.2)	14 (45.2)	14 (45.2)	26 (41.9)	14 (45.2)	46 (37.1)
High blood pressure, %	13 (41.9)	14 (45.2)	13 (41.9)	25 (40.3)	13 (41.9)	47 (37.9)
Cardiovascular disease, p-value	24 (77.4)	14 (45.2)	24 (77.4)	30 (48.4)	24 (77.4)	58 (46.8)
	0.018		0.007		0.002

* PS, propensity score; ^# Glucose-6-phosphate dehydrogenase

The results, now provided in the Supplementary Material, confirm that age is not acting as a confounding factor: the association between G6PD deficiency and increased cardiovascular risk remained statistically significant after matching. To further improve accuracy and scientific rigor, we have moderated our conclusions throughout the manuscript by replacing causal language (e.g., “clearly leads to”) with more cautious phrasing such as “is associated with,” and by expanding the Discussion of limitations accordingly. These changes strengthen the manuscript by aligning our interpretations with both the data and the constraints of the study design.

Comment 7: Acknowledge limitations more prominently and moderate clinical recommendations accordingly. The manuscript presents intriguing preliminary evidence for an important clinical association but requires additional validation and mechanistic data before definitive conclusions can be drawn about clinical management implications.

Response 7: We have incorporated the reviewer’s thoughts into the revised Conclusion (page 12, lines 462-464), emphasizing that our findings provide preliminary evidence of a potential clinical association, but that further validation and mechanistic studies are needed before definitive implications for clinical management can be established.

 

Reviewer 2 Report

Comments and Suggestions for Authors

Errigo et al present results of an observational study using historical data from 217 subjects with documented familial hypercholesterolemia (FH) and known glucose-6-phosphate dehydrogenase (G6PD) status (deficient vs not deficient). Demographic features as well as laboratory and clinical characteristics were summarized for the entire sample as well as separately by sex and G6PH status. The primary research question was whether risk of CVD was associated with G6PH status, either in a crude sense or after accounting for other known risk factors of CVD. Findings are reported through a series of tables with further elaboration and interpretation of those results in the text.

 

Overall the statistical methods are appropriate, well described, and with results presented clearly. There are issues concerning their proposed a priori power calculation/sample size justification and more minor concerns about the manner certain variables were expressed in their multivariable logistic regression. Specific statistical questions are given below with approximate line numbers from the PDF version of their submitted manuscript.

 

[Lines 189–191]. A statement is made that the power analysis was for a “two-sample t-test” and that it revealed sample size was adequate to “detect and OR as small as 1.5.” These two ideas are in conflict since the two-sample t-test is not used for assessing odds ratios. There is also some mention of “sample sizes of 217 and 9604,” the latter number having no apparent connection to the sample total size of 217 used in the study. Here’s how the authors might proceed. Since was supposed to be an a priori calculation (before seeing / collecting data), they should not be overly precise or duplicate the exact sample size(s) of groups actually obtained in the study. They report that the prevalence of G6PD deficiency (Mediterranean variant) is 10-15% [line 64], and this upper limit may be used as a guide for how the total sample size, N, would be split between G6PD deficient (15%) and proficient groups (85%). Prevalence of CVD event in a normal (non-deficient) population might be assumed to range between 30–50%. Over this range of potential prevalences, and for a total sample size of N=200 split 15%:85%, the power for a two-sided 0.05 level test to detect an odds ratio of 3.25 for CVD events is at least 78% and is over 80% provided the CVD prevalence in the normal FH population doesn’t exceed 45%. So they might report that “A priori power calculations assumed prevalence of G6PD deficiency in our sample would be at most 15% and that CVD events in the non-deficient FH population could range between 30–50%. Under these assumptions, a total sample size of N=200 (170 normal/30 deficient) patients provides at least 78.7% power to detect odds ratios of 3.25 or more at a 2-sided significance level of 5%.”

 

 

[Table 1]. The treatment of FH with statins and Ezetimibe is supported with a footnote indicating the two are not mutually exclusive, yet they still managed to produce frequency counts and percentages that sum to overall totals or to 100%, which is unexpected if things are not mutually exclusive. Would it be possible to expand the ‘Treatment’ in Table 1 to have three rows: Only statins, Only Ezetimibe, Both statins and Ezetimibe? Perhaps the counts would only appear for ‘Only statins’ and ‘Both statins and ezetimibe’ rows, and if this is the case, then it may just be a simple matter of supplying more detailed row labels.

 

[Table 3]. It appears that ‘age’ was used as a continuous covariate in the logistic regression model. Could the authors please provide some detail about the effect of age? That is, is the reported odds ratio the effect of a 1 year increase in age or something else (5 year increase?). Likewise, body mass index appears to be treated as a continuous variable in the LR model and some detail about the step size for which the OR is reported is needed (e.g., is that crude OR of 1.20 for a 1 kg/m^2 increase in BMI or for a 5 kg/m^2 increase in BMI?). Given that Table 2 shows BMI as a categorical factor with three levels, it might be more consistent if the LR model also used this same three levels, with under 25 kg/m^2 serving as the reference group and then odds ratios reported for the other two BMI levels relative to that reference.

 

Comments for author File: Comments.pdf

Author Response

Comment 1: [Lines 189-191]. A statement is made that the power analysis was for a “two-sample t-test” and that it revealed sample size was adequate to “detect and OR as small as 1.5.” These two ideas are in conflict since the two-sample t-test is not used for assessing odds ratios. There is also some mention of “sample sizes of 217 and 9604,” the latter number having no apparent connection to the sample total size of 217 used in the study. Here’s how the authors might proceed. Since was supposed to be an a priori calculation (before seeing / collecting data), they should not be overly precise or duplicate the exact sample size(s) of groups actually obtained in the study. They report that the prevalence of G6PD deficiency (Mediterranean variant) is 10-15% [line 64], and this upper limit may be used as a guide for how the total sample size, N, would be split between G6PD deficient (15%) and proficient groups (85%). Prevalence of CVD event in a normal (non-deficient) population might be assumed to range between 30–50%. Over this range of potential prevalences, and for a total sample size of N=200 split 15%:85%, the power for a two-sided 0.05 level test to detect an odds ratio of 3.25 for CVD events is at least 78% and is over 80% provided the CVD prevalence in the normal FH population doesn’t exceed 45%. So they might report that “A priori power calculations assumed prevalence of G6PD deficiency in our sample would be at most 15% and that CVD events in the non-deficient FH population could range between 30-50%. Under these assumptions, a total sample size of N=200 (170 normal/30 deficient) patients provides at least 78.7% power to detect odds ratios of 3.25 or more at a 2-sided significance level of 5%.”

Response 1: We thank the reviewer for highlighting the inaccuracies and inconsistencies in our original description of the power analysis. We agree that the initial wording was misleading and potentially confusing. In the revised manuscript, we have rewritten this section entirely following the reviewer’s recommendations (page 5, lines 190–195).

Comment 2: [Table 1]. The treatment of FH with statins and Ezetimibe is supported with a footnote indicating the two are not mutually exclusive, yet they still managed to produce frequency counts and percentages that sum to overall totals or to 100%, which is unexpected if things are not mutually exclusive. Would it be possible to expand the ‘Treatment’ in Table 1 to have three rows: Only statins, Only Ezetimibe, Both statins and Ezetimibe? Perhaps the counts would only appear for ‘Only statins’ and ‘Both statins and ezetimibe’ rows, and if this is the case, then it may just be a simple matter of supplying more detailed row labels.

Response 2: As no patients in our cohort were treated with ezetimibe alone, we have revised Table 1 accordingly. The treatment categories are now reported as “Only statins” and “Statins plus ezetimibe,” which resolves the inconsistency and clarifies the frequency counts. This correction has been incorporated into the revised manuscript.

Comment 3: [Table 3]. It appears that ‘age’ was used as a continuous covariate in the logistic regression model. Could the authors please provide some detail about the effect of age? That is, is the reported odds ratio the effect of a 1 year increase in age or something else (5 year increase?). Likewise, body mass index appears to be treated as a continuous variable in the LR model and some detail about the step size for which the OR is reported is needed (e.g., is that crude OR of 1.20 for a 1 kg/m^2 increase in BMI or for a 5 kg/m^2 increase in BMI?). Given that Table 2 shows BMI as a categorical factor with three levels, it might be more consistent if the LR model also used this same three levels, with under 25 kg/m^2 serving as the reference group and then odds ratios reported for the other two BMI levels relative to that reference.

Response 3: We thank the reviewer for these important observations. We have clarified in the table footnotes that the odds ratios for age correspond to a one-year increase (page 8, lines 285-286). To improve consistency and interpretability, we re-ran the logistic regression treating BMI as a categorical variable, in line with its presentation in Table 2. In the revised manuscript, the Results section and updated Table 2 now report odds ratios for BMI categories, with <25 kg/m² as the reference group. The table legend also specifies the step size for age, ensuring clarity in the interpretation of effect estimates.

In summary, we have carefully addressed each point. The corrected text, additional analyses, and clarifications have been incorporated into the revised manuscript. We believe these changes have significantly improved the clarity and accuracy of our work. We thank the reviewer again for his/her valuable feedback.

 

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

n/a

Author Response

Comment 1: n/a

 

Response 1: We would like to thank the reviewer for his/her appreciation.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for attending to the statistical concerns. The manuscript is, I think, more clear.

The new material involving propensity score matching raises some new points that weren't in the previous version.

1. Propensity score matching is most often seen in observational studies where the "risk factor" is something that could have been randomly assigned by a researcher (if there were no ethical barriers, as in the example of randomly assigning people to a smoking vs non-smoking group). PSM is not often used for matching on the basis of something that could not really be assigned: developing the matching model for one's propensity to be male vs female or to have (or not have) a genetic trait would be examples where this sort of random allocation can't happen and where PSM isn't the best option.
2. PSM is also something done usually when there are large amounts of data, both for the purposes of building that initial logistic regression model that estimates the "propensity" on which matching is then conducted, and also to make certain that the final matched sets are of sufficient size to perform the analysis of interest. 
3. Showing results of 1:1, 1:2, and 1:4 matching is something of an overkill, and at the highest extreme (1:4 matching) produces results that aren't much of an improvement over the initial sample sizes. All analyses use the same n=31, but just increase size of the 'control' group, which approaches the full number of controls at the highest matching ratio (from 124 matched vs 186 full and unmatched). It would have been cleaner to just stick with the 1:1 matching.
4. There is an idea of "double robustness" in PSM whereby the analysis of the matched data also includes the same variables used in creation of the propensity score weights. That is, the analysis of cardiovascular disease (the response) for the matched data would include sex, age, bmi, smoking status, and high blood pressure indicators as well. This could be valuable in situations like yours where the sample size(s) are smaller than what is typically used.
5. The average age in the propensity score set (Table S4) is 57.9 +/- 15.8 years, but in the main manuscript it is reported as 58.2 +/- 15.9 years (Table 2) for the same cohort of n=31. Could you please re-check and update as needed?

Author Response

Comment 1: Thank you for attending to the statistical concerns. The manuscript is, I think, more clear.

Response 1: We sincerely thank the reviewer for the constructive feedback on our revised manuscript

 

Comment 2: The new material involving propensity score matching raises some new points that weren't in the previous version.

Propensity score matching is most often seen in observational studies where the "risk factor" is something that could have been randomly assigned by a researcher (if there were no ethical barriers, as in the example of randomly assigning people to a smoking vs non-smoking group). PSM is not often used for matching on the basis of something that could not really be assigned: developing the matching model for one's propensity to be male vs female or to have (or not have) a genetic trait would be examples where this sort of random allocation can't happen and where PSM isn't the best option.

Response 2: We greatly appreciate the clarification regarding the appropriate contexts for the use of propensity score matching (PSM), and we agree with the reviewer’s perspective that PSM is more commonly applied to exposures that could, in principle, be randomly assigned. We acknowledge that G6PD deficiency, as a genetic trait, is inherently different from such a scenario. Our rationale for applying PSM was primarily to address the reviewer’s initial concern about age imbalance and to demonstrate that our findings are not explained by this factor. That's why we've confined it to the supplementary material.

 

Comment 3: PSM is also something done usually when there are large amounts of data, both for the purposes of building that initial logistic regression model that estimates the "propensity" on which matching is then conducted, and also to make certain that the final matched sets are of sufficient size to perform the analysis of interest.

Showing results of 1:1, 1:2, and 1:4 matching is something of an overkill, and at the highest extreme (1:4 matching) produces results that aren't much of an improvement over the initial sample sizes. All analyses use the same n=31, but just increase size of the 'control' group, which approaches the full number of controls at the highest matching ratio (from 124 matched vs 186 full and unmatched). It would have been cleaner to just stick with the 1:1 matching.

 

Response 3: Following the reviewer’s advice, we have revised the manuscript and we have simplified our supplementary material to present only the 1:1 matching results, which we agree are the clearest and most appropriate for our dataset. This adjustment avoids redundancy while still addressing covariates as potential confounders.

 

Comment 4: There is an idea of "double robustness" in PSM whereby the analysis of the matched data also includes the same variables used in creation of the propensity score weights. That is, the analysis of cardiovascular disease (the response) for the matched data would include sex, age, bmi, smoking status, and high blood pressure indicators as well. This could be valuable in situations like yours where the sample size(s) are smaller than what is typically used.

Response 4: See above

 

Comment 5: The average age in the propensity score set (Table S4) is 57.9 +/- 15.8 years, but in the main manuscript it is reported as 58.2 +/- 15.9 years (Table 2) for the same cohort of n=31. Could you please re-check and update as needed?

Response 5 : Thank you for noting the discrepancy in the average age reported for the G6PD-deficient subgroup (Table S4 vs. Table 2). In the revised  manuscript the discrepancy was corrected (Table 2 at page 7).

Finally, we have added a sentence in the Limitations section to explicitly acknowledge that PSM may not be the optimal approach for genetic traits and that our use of it was intended to provide reassurance about confounding rather than to suggest causal inference (page 12, lines 436-438).

We are grateful to the reviewer for these insightful suggestions, which have led to meaningful improvements in the clarity, rigor, and transparency of our work.