Factors Associated with Secondary Pulmonary Hypertension Among Hospitalized Females: An Artificial Neural Network Analysis of a National US Cohort

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

I read with interest the manuscript by Mohammed and colleagues, which investigates predictors associated with hospitalization for secondary pulmonary hypertension among women using the 2019 National Inpatient Sample and an artificial neural network model.

The very large sample size and the attempt to apply a machine-learning approach are strengths of the study. The focus on female patients is an added value. 

However, in its current form, the paper needs to address several points before it can be considered for publication. Here my suggestions:

• Since secondary pulmonary hypertension may derive from left heart disease, lung disease, chronic thromboembolic disease, or multifactorial mechanisms, the authors should better acknowledge the risk of misclassification and avoid treating SPH as a single uniform clinical entity.
• The manuscript describes predictors of hospitalization, but the analysis seems to compare hospitalized women with and without coded SPH within the NIS database. Therefore, the model appears to predict the presence of an SPH diagnosis among hospitalized women, rather than future hospitalization risk among patients with known SPH. This distinction should be better clarified throughout the abstract, methods, results, and conclusions.
• An AUC of 0.823 is encouraging, but the manuscript should not imply that the ANN can already function as a clinically actionable risk tool. The authors should clearly state that the model needs to be further validated.
• The methods section states that ANNs offer several advantages over multivariable regression, but no direct comparison with a conventional logistic regression model is provided. Since logistic regression remains highly interpretable and widely used in clinical epidemiology, the authors should either include a comparative model or soften this statement.
• The authors should discuss whether the ANN is identifying disease-specific predictors or simply capturing overall frailty, comorbidity burden, and healthcare utilization.
• Some pathophysiological explanations, including estrogen-related effects, autoimmune burden, and environmental exposures, are interesting but remain speculative in the context of this administrative dataset.
• The authors should specify whether the proposed use is screening, coding-based risk stratification, identification of high-comorbidity hospitalized patients, or development of future predictive tools. This would improve the translational value of the paper.

Author Response

We express our sincere gratitude to Reviewer 2 for acknowledging the strengths of the study and for providing a comprehensive set of comments that have significantly enhanced its rigor and clarity.

Comment 2.1: Since secondary pulmonary hypertension may derive from left heart disease, lung disease, chronic thromboembolic disease, or multifactorial mechanisms, the authors should better acknowledge the risk of misclassification and avoid treating SPH as a single uniform clinical entity.

Response: Thank you for pointing this out. The points have been acknowledged and addressed in accordance with Comment 1.2. The abstract and introduction now describe SPH as a heterogeneous entity extending across groups 2–5, and the Limitations section elucidates the risk of misclassification. It is explicitly stated that SPH should not be considered as a single, uniform clinical entity. (Page number:2, 9, Line number:72-78; 288-296)

Comment 2.2: The manuscript describes predictors of hospitalization, but the analysis seems to compare hospitalized women with and without coded SPH within the NIS database. Therefore, the model appears to predict the presence of an SPH diagnosis among hospitalized women, rather than future hospitalization risk among patients with known SPH. This distinction should be better clarified throughout the abstract, methods, results, and conclusions.

Response: Thank you for pointing this out. Agreed; this has now been corrected uniformly throughout the abstract, methods, results, and conclusions (see Comment 1.1).

Comment 2.3: An AUC of 0.823 is encouraging, but the manuscript should not imply that the ANN can already function as a clinically actionable risk tool. The authors should clearly state that the model needs to be further validated.

Response: Thank you for your suggestion. We now explicitly clarify that the model is neither a validated screening instrument nor an actionable risk assessment tool. Furthermore, an AUC of 0.823 alone does not constitute sufficient evidence of clinical utility. Additional external and prospective validation is necessary. Such statements are documented in the abstract, discussion, and conclusions. (Page number:9, Line number:298-300, 306-309)

Comment 2.4: The methods section states that ANNs offer several advantages over multivariable regression, but no direct comparison with a conventional logistic regression model is provided. Since logistic regression remains highly interpretable and widely used in clinical epidemiology, the authors should either include a comparative model or soften this statement.

Response: Thank you for your suggestion. We have softened the claim of ANN's superiority to present a balanced statement that recognizes logistic regression as highly interpretable and broadly utilized, while noting that the advantages of ANN are contingent upon specific contexts. Additionally, we have included, as a limitation and avenue for future research, the absence of any comparison with logistic regression. (Page number:8, 9, Line number:267-271, 298-300)

Comment 2.5: The authors should discuss whether the ANN is identifying disease-specific predictors or simply capturing overall frailty, comorbidity burden, and healthcare utilization.

Response: Thank you for your suggestion. We have included a paragraph acknowledging that several of the strongest correlates (age, CKD, COPD, complicated hypertension) are also general markers of multimorbidity, frailty, and healthcare utilization. Additionally, it is recognized that the model may partially reflect an older, multimorbid inpatient phenotype rather than SPH-specific biology. (Page number:8, Line number:220-226)

Comment 2.6: Some pathophysiological explanations, including estrogen-related effects, autoimmune burden, and environmental exposures, are interesting but remain speculative in the context of this administrative dataset.

Response: Thank you for your suggestion. We now explicitly frame these as mechanistic hypotheses derived from the extensive literature, which cannot be empirically tested within this administrative dataset, and present them as contextual considerations rather than results of this study. (Page number:7, Line number:189-203)

Comment 2.7: The authors should specify whether the proposed use is screening, coding-based risk stratification, identification of high-comorbidity hospitalized patients, or development of future predictive tools. This would improve the translational value of the paper.

Response: Thank you for your suggestion. We have incorporated an “intended use” statement clarifying that the findings are hypothesis-generating and may, upon validation, contribute to the development of personalized risk stratification tools or the identification of high-comorbidity inpatients for more detailed assessment. However, these findings are not intended to serve as a screening instrument or a pre-existing risk calculator. (Page number:8, Line number:264-267)

Reviewer 2 Report

Comments and Suggestions for Authors

The Authors propose an interesting manuscript. The manuscript is valuable because it provides insights resulting from querying a large database using ANN, which is capable of detecting subtle patterns.

 

However, the manuscript requires a series of revisions to be clear and reliable.

I have formulated a series of recommendations/observations.

 

The structure of an JPM article usually do not begin with Study Highlights. This section can be moved later in the manuscript.

Introduction

Introduction needs to be expanded. Some suggestions (but not limited to):

Research shows an increased prevalence of SPH among females – Authors are asked to bring some data from the existing literature.

However, comprehensive large-scale data on predictors of hospitalizations in women with PH is available[2] – Major findings should be highlighted.

Methods

Lines 102-103: Following our research objectives, variables that are capable of predicting SPH were identified through a comprehensive literature review. Comment: How exactly was the selection made (how were the significant variables, and therefore of interest, identified)?

Lines 108-109: ... prior history of Myocardial Infarction (MI) with prior revascularization...

Comment: Please consider „history of myocardial infarction (MI) with revascularization”

Lines 106- 109: Additionally, we considered comorbid medical conditions, including ... congenital heart disease...

Comment: Pulmonary arterial hypertension (PH group 1) includes among etiologies congenital heart disease with a left to right shunt. Therefore, congenital heart disease should be an exclusion and not an inclusion criterion.

Were exclusion criteria considered at all?

 

Results

Figure 1. The text of the figure is extremely difficult to read. Authors are asked to optimize the quality of the figure.

Table 1

Comment: Two parameters were used as exponents of atherosclerotic coronary disease, namely „Prior Myocardial Infarction” and „Prior CABG”. Why there is no "prior elective stenting" category? In many countries, patients with obstructive CAD are identified and managed with early intervention (stenting), before an acute coronary event.

What is complicated HTN? An explanatory phrase is necessary.

What includes „arthropathies”? The etiological spectrum is wide and some etiologies are associated with increased cardiovascular risk or with concomitant microvascular damage. An explanatory phrase is necessary.

The list of relative importance of predictors stops at AIDS (22.6%). Why this threshold (higher than 20%)?

Discussion

Line 160: SPH is more prevalent among older females...Comment: It would be useful to add age. It would give the results a better clinical impact.

Line 163: The authors start the enumeration of mechanisms with firstly, but then there is no secondly and so on.

Lines 167-168. autoimmune disorders in females and exposure to fetal cells, may elevate the risk of developing SPH. Comment:  a reference and a brief explanation of the mechanisms are needed.

Lines 172-173: A study by Ventetuolo et al. has demonstrated that males with PH experience worse survival outcomes compared to females [8]. This statement contradicts an early statement (in Introduction, lines 85-86): A decade of surveillance from 2001 to 2010 demonstrates an increase in female mortality and hospitalizations relative to males [2]. The results of the present study should be commented taking into account these discrepant results of the literature and, in addition, the causes of the discrepancies should be commented on.

Lines 175-176: An increased prevalence of cardiovascular and atherosclerotic cardiovascular disease (ASCVD) risk factors... Comment: prior MI with revascularization is not a risk factor. It belongs to MACE. Therefore, this statement must be reconsidered and written correctly.

In general, the Discussion section lacks explanations (mechanisms, pathophysiological pathways) of the observed associations.

Thank you!

Author Response

We express our gratitude to Reviewer 3 for their comprehensive, detailed review, which has significantly enhanced the precision and clarity of the manuscript.

Comment 3.1: The structure of an JPM article usually do not begin with Study Highlights. This section can be moved later in the manuscript.

Response: Thank you for pointing this out. The Study Highlights have been removed from the manuscript.

Comment 3.2: Introduction needs to be expanded. Some suggestions (but not limited to): Research shows an increased prevalence of SPH among females – Authors are asked to bring some data from the existing literature. However, comprehensive large-scale data on predictors of hospitalizations in women with PH is available[2] – Major findings should be highlighted.

Response: Thank you for the comment. The Introduction has been expanded to include the surveillance findings of George et al. [2], which indicate that PH-related hospitalization and mortality rates have consistently been higher among women than men during the period from 2001 to 2010, with the disparity widening over time. Additionally, a corrected statement has been incorporated to clarify that comprehensive data on factors associated with SPH among hospitalized women remain limited. It is noted that the original sentence, which claimed such data “is available," was a typographical error that contradicted the rationale of the study; this has been amended to accurately state that such data "remain limited.” (Page number:2, Line number:79-82, 84)

Comment 3.3: How exactly was the selection made (how were the significant variables, and therefore of interest, identified)?

Response: Thank you for the comment. The Methods section now indicates that candidate predictors were predetermined based on a comprehensive literature review, clinical plausibility, and their availability within the NIS data elements. (Page number:3, Line number:113-124)

Comment 3.4: Please consider „history of myocardial infarction (MI) with revascularization”

Response: Thank you for your suggestion. We have now rephrased the phrase as recommended to “history of myocardial infarction (MI) with revascularization” throughout the manuscript. (Page number:3, Line number:122)

Comment 3.5: Pulmonary arterial hypertension (PH group 1) includes among etiologies congenital heart disease with a left to right shunt. Therefore, congenital heart disease should be an exclusion and not an inclusion criterion. Were exclusion criteria considered at all?

Response: Thank you for pointing this out. We agree. Congenital heart disease has been excluded from the list of candidate predictors and is no longer regarded as a comorbidity of interest. Additionally, we have incorporated a statement reflecting this change. An inclusion/exclusion criteria paragraph has been added, and a Limitation section acknowledges that administrative coding cannot definitively exclude Group 1 PAH. (Page number:3, Line number:103-112)

Comment 3.6: Figure 1. The text of the figure is extremely difficult to read. Authors are asked to optimize the quality of the figure.

Response: Thank you for your suggestion. We replaced Figure 1 with a high-resolution version featuring enlarged, legible axis labels and predictor names; the figure file is included with the resubmission. (Page number:7)

Comment 3.7: Two parameters were used as exponents of atherosclerotic coronary disease, namely „Prior Myocardial Infarction” and „Prior CABG”. Why there is no "prior elective stenting" category? In many countries, patients with obstructive CAD are identified and managed with early intervention (stenting), before an acute coronary event.

Response: Thank you for the comment. Prior percutaneous coronary intervention (PCI) has been added to Table 1 (0.7% in the SPH group vs 0.3% in the non-SPH group; p<0.001). Like prior CABG, it is reported descriptively; it was not among the variables entered into the neural network. (Page number:5)

Comment 3.8: What is complicated HTN? An explanatory phrase is necessary.

Response: Thank you for the comment. A footnote in Table 1 now delineates complicated hypertension (hypertension with documented end-organ involvement) from uncomplicated hypertension (essential hypertension without coded target-organ damage), in accordance with the AHRQ/Elixhauser classification. (Page number:6)

Comment 3.9: What includes “arthropathies”? The etiological spectrum is wide and some etiologies are associated with increased cardiovascular risk or with concomitant microvascular damage. An explanatory phrase is necessary.

Response: Thank you for the comment. A footnote in Table 1 now delineates the arthropathies category (inflammatory and non-inflammatory joint disorders) and emphasizes that the inflammatory/connective-tissue subset is of particular relevance to pulmonary vascular disease. Comorbidities including arthropathies were defined from ICD-10-CM codes using AHRQ Clinical Classifications groupings (Page number:6)

Comment 3.10: The list of relative importance of predictors stops at AIDS (22.6%). Why this threshold (higher than 20%)?

Response: Thank you for pointing this out. We now clearly specify the rationale. Variables with normalized importance over 20% are marked as the main contributors, and the full ranking extends to the complete ranking of all 23 predictors down to diabetes without chronic complications (2.9%). In addition, HIV/AIDS has been added to Table 1 (0.43% vs 0.36%; p<0.001), so every ranked predictor also appears in the baseline table. (Page number:5, 6, Line number:175-180)

Comment 3.11: It would be useful to add age. It would give the results a better clinical impact.

Response: Thank you for your suggestion. Age has been included: “median age 75 vs. 58 years.” (Page number:7, Line number:186)

Comment 3.12: The authors start the enumeration of mechanisms with firstly, but then there is no secondly and so on.

Response: Thank you for pointing this out. The enumeration of mechanisms has been corrected to incorporate First, Second, Third, and Fourth. (Page number:7, Line number:191-204)

Comment 3.13: autoimmune disorders in females and exposure to fetal cells, may elevate the risk of developing SPH. Comment:  a reference and a brief explanation of the mechanisms are needed.

Response: Thank you for your suggestion. We have provided a succinct mechanistic explanation, which includes the female predominance of autoimmune and connective-tissue diseases (notably associated with pulmonary vascular disease) and fetal microchimerism, referring to the persistence of fetal cells in the maternal circulation, implicated in autoimmune vascular injury now supported by a dedicated reference (Nelson & Lambert, Semin Immunopathol 2025; reference [6]). (Page number:7, Line number:199-204)

Comment 3.14: A study by Ventetuolo et al. has demonstrated that males with PH experience worse survival outcomes compared to females [8]. This statement contradicts an early statement (in Introduction, lines 85-86): A decade of surveillance from 2001 to 2010 demonstrates an increase in female mortality and hospitalizations relative to males [2]. The results of the present study should be commented taking into account these discrepant results of the literature and, in addition, the causes of the discrepancies should be commented on.

Response: Thank you for pointing this out. We have included a paragraph that reconciles the two perspectives. “Population-level surveillance indicates that women bear a greater absolute burden of PH-related hospitalizations and mortality [2], reflecting their higher prevalence of PH. Conversely, within diagnosed pulmonary arterial hypertension cohorts, male sex is associated with poorer survival, often referred to as the ‘estrogen paradox’ as reported by Ventetuolo et al. [9] and in more recent reviews [4,5]. These findings are not mutually exclusive; rather, they describe different metrics (population burden versus case-fatality rates) across distinct populations (all-cause or secondary PH versus group 1 PAH)…” (Page number:7, 8, Line number:209-219)

Comment 3.15: Prior MI with revascularization is not a risk factor. It belongs to MACE. Therefore, this statement must be reconsidered and written correctly.

Response: Thank you for pointing this out. Risk factors, including hypertension, hyperlipidemia, obesity, and diabetes, are now delineated separately from documented atherosclerotic disease and its complications, such as prior myocardial infarction (MI), revascularization, and peripheral vascular disease. An explicit clarification is provided that prior MI and revascularization are indicative of established cardiovascular disease or major adverse cardiovascular events rather than merely risk factors. (Page number:8, Line number:227-233)

Comment 3.16: In general, the Discussion section lacks explanations (mechanisms, pathophysiological pathways) of the observed associations.

Response: Thank you for your suggestion. We strengthened the mechanistic content by incorporating the connection between arthropathies and connective tissue disorders, as well as the sections on reconciliation and comorbidity burden, in addition to the existing mechanistic explanations for COPD, prior VTE, CKD, diabetes, and obesity. (Page number:8, Line number:251-253)

Please see the attachment for responses to all the reviewers.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript asserts that it identifies predictors for hospitalization with secondary pulmonary hypertension. However, the study only uncovers characteristics linked to the diagnosis of SPH among hospitalized female patients already admitted. The important distinction should be consistently included in the title, abstract, discussion, and conclusions.

-  The authors consider ICD-10 code I27.2 to represent "non-group 1 pulmonary hypertension" and use this term for the entirety of the manuscript.

Given the heterogeneity of pulmonary hypertension groups 2–5, the authors should discuss in  detail: the validity of using ICD coding to define SPH, potential misclassification bias, the inability to distinguish specific PH groups within the NIS database.

This limitation may substantially influence interpretation of the findings.

• check again the definition of pulmonary hypertension, according to the recent guidelines.
• ANN methodology is insufficiently described. Please provide more details.
• From a methodological point of view, the reader finds out the characteristics of the women admitted with diagnostic code I27.2, and not the hospitalization predictors. All these patients were already admitted. Please provide more explanations.
• Interpretation of  these predictors should be more cautious
• what you mean by:   women experience "a more severe form of the disease" ? What data did you consider?

Author Response

We thank Reviewer 1 for identifying the central framing issue and several important methodological points.

Comment 1.1:  The study only uncovers characteristics linked to the diagnosis of SPH among hospitalized female patients already admitted. The important distinction should be consistently included in the title, abstract, discussion, and conclusions.

Response: We fully agree and have corrected this throughout. The title is now “Factors Associated with Secondary Pulmonary Hypertension Among Hospitalized Females …” The abstract, methods, results, discussion, and conclusions consistently state that we identify factors associated with a coded SPH diagnosis among hospitalized women, not predictors of hospitalization. For example, the conclusion now reads that the model “classifies an existing diagnosis rather than predicting future hospitalization.” (Page numbers:1, 2, 10, Line number:2-4; 96-98, 305-309)

Comment 1.2: The authors consider ICD-10 code I27.2 to represent "non-group 1 pulmonary hypertension" and use this term for the entirety of the manuscript. Given the heterogeneity of pulmonary hypertension groups 2–5, the authors should discuss in detail: the validity of using ICD coding to define SPH, potential misclassification bias, the inability to distinguish specific PH groups within the NIS database. This limitation may substantially influence interpretation of the findings.

Response: Thank you for pointing this out. We have included a comprehensive discussion. The Methods section no longer characterizes the I27.2x code as “validated” for SPH. The Limitations now specify that I27.2x has not undergone formal validation for SPH, is subject to misclassification, encompasses PH groups 2–5, and that the NIS cannot reliably differentiate these subgroups. We caution that this heterogeneity may significantly impact interpretation and advise that SPH should not be regarded as a singular, uniform entity. (Page number:9, Line number:288-293)

Comment 1.3: Check again the definition of pulmonary hypertension, according to the recent guidelines.

Response: Corrected. PH is now defined by a resting mean pulmonary arterial pressure exceeding 20 mmHg, in accordance with the consensus established by the 2022 European Society of Cardiology/European Respiratory Society (ESC/ERS) guidelines. This threshold marking the definition has been revised downward from the previous standard of ≥25 mmHg. The 2022 guideline has been added as a reference [1]. (Page number:2, Line number:67-71)

Comment 1.4: ANN methodology is insufficiently described. Please provide more details.

Response: Thank you for the comment. The Methods section now delineates the architecture of a multilayer perceptron in IBM SPSS v25.0 with a 53-unit input layer (23 candidate variables after coding), one hidden layer of 9 units with hyperbolic-tangent activation, and a two-unit softmax output layer with cross-entropy error; 70:30 split; percent incorrect predictions and AUC; normalized variable importance; internal validation only. Additionally, the 17,236,228 is the weighted national estimate, whereas the ANN used the unweighted analytic sample of 3,319,543 (127,703 of 3,447,246 excluded for missing data. (Page number:3, Line number:125-147)

Comment 1.5: From a methodological point of view, the reader finds out the characteristics of the women admitted with diagnostic code I27.2, and not the hospitalization predictors. All these patients were already admitted. Please provide more explanations.

Response: Thank you for the comment. This is addressed through the reframing in Comment 1.1, by an additional Methods statement elucidating the encounter-level, cross-sectional comparison of hospitalized women with and without the code, and by a Limitations section noting that the design does not determine future hospitalization risk. (Page number:9, Line number:287-293)

Comment 1.6: Interpretation of these predictors should be more cautious.

Response: Thank you for your suggestion. Interpretation is now more cautious throughout. We describe the variables as factors associated with a coded diagnosis; additionally, a paragraph is included noting that several may reflect general comorbidity burden, frailty, and healthcare utilization. The term “excellent” is replaced with “good discrimination,” and it is stated that the model is hypothesis-generating and not clinically actionable without validation. (Page number:6, Line number:168)

Comment 1.7:  What is meant by women experiencing “a more severe form of the disease”? What data did you consider?

Response: Thank you for the comment. We have excluded this characterization from our report. Our administrative data do not encompass measures of disease severity, and the claim was not substantiated by the cited source. The revised sentence now solely presents the treatment-pattern findings referenced. [18] (women more likely to receive medical rather than surgical therapy in a CTEPH referral cohort). (Page number:9, Line number:276-278)

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript is substantially improved and has addressed most of my previous concerns.

I would only suggest reporting additional model performance metrics, such as sensitivity, specificity, predictive values, calibration, and/or a confusion matrix. Minor polishing of tables, figures, and language would further improve the manuscript.

Author Response

Thank you, please find the attached response letter.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

dear authors, thank you for assessing all the issues from the review report.

Author Response

Thank you.

Author Response File: Author Response.pdf