Predicting the Risk of Alzheimer’s Disease and Related Dementia in Patients with Mild Cognitive Impairment Using a Semi-Competing Risk Approach

Round 1

Reviewer 1 Report

The research topic and approach are interesting

Author Response

Thank you for your comments.

Reviewer 2 Report

In the manuscript entitled “Predicting risk of Alzheimer's disease and related dementia in 2 patients with mild cognitive impairment using a semi-competing risk approach”, Chen et al. proposed a modified model from mild cognitive impairment (MCI) to predict the risk of Alzheimer’s disease (AD) disease. A semi-competing risk approach was introduced into this model, which can take competing risks of death into consideration. The comments are given as below.

1.      With this model, it is failed to obtain a significant improvement compared with the traditional method. Though the analyzed results showed a bit difference in some factors such as renal disease and liver disease, the authors did not explain it in physiological or pathological way.

2.      The manuscript used conditional probability function to revise the prediction model. However, with no solid proof, we can hardly treat death and AD as separated events. In fact, AD brings complicated physiological change into body system, and of cause associates with death risk.

3.      The manuscript needs carefully check for grammatical errors and inconsistencies.

Author Response

1. With this model, it is failed to obtain a significant improvement compared with the traditional method. Though the analyzed results showed a bit difference in some factors such as renal disease and liver disease, the authors did not explain it in physiological or pathological way.

 

Thanks for the comments.  We have added discussions on factors such as renal disease, and how to interpret the modelling results.

 

2. The manuscript used conditional probability function to revise the prediction model. However, with no solid proof, we can hardly treat death and AD as separated events. In fact, AD brings complicated physiological change into body system, and of cause associates with death risk.

 

Thanks for the comments.  This is exactly the reason the choice of an appropriate model is important (i.e., death and AD are not separated events, but related).  The semi-competing risk model is the appropriate model for the joint risk of the two events, rather than ignoring the dependence between events.  Our current paper focuses on AD/ADRD as the final outcome and as when death occurred before AD/ADRD is a competing risk for the progression from MCI to AD. 

 

Death after AD/ADRD or caused by ADRD are not currently considered, but is actually another interesting research question to explore.  Nevertheless, EHR data often do not contain cause of death (e.g., whether the death was attributed to AD), especially death that occurred outside of the hospital.  We did link the OneFlorida+ EHRs with national death index, so that we can accurately know when the patient died, but not what caused the death.  Without accurate cause of death, it will introduce significant biases into the model.  Nevertheless, a good future work.

 

3. The manuscript needs carefully check for grammatical errors and inconsistencies.

 

Thanks for the comments.  We have carefully checked and made changes where applicable.

 

Reviewer 3 Report

This paper considers the rate of progression from MCI to AD/ADRD, while treating death as a competing risk under the semi-competing risk framework. The authors propose cause-specific proportional hazard models for the time from MCI to AD/ADRD, time from AD/ADRD to death, and time to death when AD/ADRD is not observed before death. Specifically, the authors propose a shared frailty model to be estimated under a Bayesian framework. The authors then apply model and estimation methods to the OneFlorida+ network data and investigate hazard ratios of various risk factors. I have a few comments.

 

1. It is known that if death is treated as censoring in the estimation of hazard ratios using the standard coxph method and packages (later referred to as the "standard method"), the obtained hazard ratios are for the cause-specific hazard and there is no bias involved. The authors seem to have failed to mention or recognize this throughout this paper, which created a few issues:

- This important way of estimating cause-specific HR under proper competing risk assumptions were not discussed in the Introduction;

- In table 2, the authors seem to have made the mistake of referring to the correct "standard method" as the method that ignores death as a competing risk. This is not true. The difference between the "standard method" and the authors' method is that the "standard method" has a "marginal" interpretation (not conditional on frailty), while the authors' method has a "conditional" interpretation (conditional on frailty).

- Connected to the above point, the interpretation of model coefficients in the Discussion (or maybe elsewhere) is incorrect. The explanation for having similar results from two methods is also a big too strong, all things considered.

- I would suggest the authors either include these important discussions on the "standard method", or clarify in the case of me having a misunderstanding.

2. The introduction of notations $T_1$ and $T_2$ in Section 2 is incorrect (reversed). It is also a bit unclear what the hazard ratio model assumptions are for, specifically for which variable under which scenario.

3. I might have missed it, but it is unclear what distribution is used for the shared frailty.

4. Given 1, it is also important for the authors to better elaborate why the authors' method is favorable in comparison with the standard method.

Author Response

 - I would suggest the authors either include these important discussions on the "standard method", or clarify in the case of me having a misunderstanding.

Thanks for the comment.  As we described in the manuscript, the rational of using semi-competing approach is that “In the elderly population whose first MCI diagnosis older than 65 years, patients are subject to both the risk of AD/ADRD and the risk of death because of age. The competing risk of death would censor the AD/ADRD outcomes, thus serves as an informative censoring for AD/ADRD failure events.”  We have also added more discussions on the comparison between the two methods in the discussion section.

-The introduction of notations $T_1$ and $T_2$ in Section 2 is incorrect (reversed). It is also a bit unclear what the hazard ratio model assumptions are for, specifically for which variable under which scenario.

 

Thank you for your review and comments on our paper.  We appreciate your feedback and would like to address your concerns regarding the introduction of notations $T_1$ and $T_2$ in Section 2.  We apologize for any confusion caused by the reversal of notations $T_1$ and $T_2$ in the section, and we have revised the manuscript accordingly to ensure clarity.

 

In addition, we have provided a more detailed explanation of the hazard ratio model assumptions for each variable and scenario, to help readers better understand our methodology.  We hope that these revisions will address your concerns and improve the clarity of our manuscript.  Once again, thank you for your comments, and please let us know if you have any further questions or feedback.

 

- I might have missed it, but it is unclear what distribution is used for the shared frailty.

Thank you for your review and comments on our manuscript. We appreciate your feedback and would like to address your question regarding the distribution used for the shared frailty.

We assumed that the random effect or frailty follows a gamma distribution, which is a common distribution used for modeling random effects. This gamma distribution assumption is based on the assumption that the individual frailties are non-negative and have a skewed distribution, which is a common characteristic of frailties in survival analysis. We hope that this revised explanation clarifies the distribution used for the shared frailty in the semi-competing risks model in our study. We appreciate the reviewer's attention to detail in pointing out the error. If you have any further questions or concerns, please do not hesitate to let us know.

 

- Given 1, it is also important for the authors to better elaborate why the authors' method is favorable in comparison with the standard method.

Thanks for the comments, we have added more discussions on the comparison between the two methods in the discussion section.

 

Reviewer 4 Report

Thanks for submitting and taking the time to develop this piece. It was fairly organized and linked together. There was a reasonable flow of thought in the manuscript. A sound methodology was employed in the study, which highlighted its novelty. In addition to the detailed explanations of the results and discussion rounds, substantial references are included to support the points made. In the future, the study will likely take a different direction. It is necessary to pay more attention to English grammar and sentence construction in order to improve content quality. Thanks.

A semi-competing risk approach can be used to predict the risk of Alzheimer's disease and related dementia in patients with mild cognitive impairment. This approach models the competing risk of death alongside the risk of developing dementia, which is important given that death may occur before dementia onset. Here are some possible areas of strength and weakness of this study:

  Strengths:   By considering the competing risk of death, the approach provides a more accurate estimate of the risk of dementia compared to models that ignore death as a possible outcome. The model parameters also were interpreted, making it easier to understand the factors that contribute to the risk of dementia, while the sample size of patients with mild cognitive impairment was adequate, such that it increases the accuracy of the model predictions.   Weaknesses:   The semi-competing risk approach assumes that the risk of dementia and the risk of death are independent. This may not be true in all cases, as factors that increase the risk of dementia may also increase the risk of death. Also, there are unmeasured confounding variables that can bias the results of the model, leading to inaccurate predictions. All these need to be accounted for in the study.

Author Response

Thanks for the comments.  We have acknowledged and added unmeasured confounding as one of the limitations in the paper.

In addition, it is important to clarify that the semi-competing risk approach we used in our paper explicitly accounts for the fact that the risk of dementia and death are not independent, but rather are correlated.

We acknowledge that there may be some cases where factors that increase the risk of dementia also increase the risk of death. However, the semi-competing risk approach we used is specifically designed to handle such situations and can provide valuable insights into the joint risk of these two events.

Overall, we believe that the semi-competing risk approach is a valuable tool for analyzing complex data and can help researchers to better understand the relationship between different health outcomes. We appreciate the reviewer's feedback and will make sure to clarify these points in the revised version of our manuscript.

Round 2

Reviewer 2 Report

 

Thanks for the response of the authors. In the manuscript entitled “Predicting the risk of Alzheimer's disease and related dementia in patients with mild cognitive impairment using a semi-competing risk approach”, the authors made corresponding amendments to the questions. But there are still some questions in details:

1.      I still have concerns about whether this model bring more information than the conventional way. Please provide more evidence on a certain clinical case.

 

2.      The second question is about the precision and sensitivity of the model on prediction of the risk of AD. Is this model reliable to predict AD risk? Please provide more details in the paper.

 

 

Author Response

Thanks for the response of the authors. In the manuscript entitled “Predicting the risk of Alzheimer's disease and related dementia in patients with mild cognitive impairment using a semi-competing risk approach”, the authors made corresponding amendments to the questions. But there are still some questions in details:

1. I still have concerns about whether this model bring more information than the conventional way. Please provide more evidence on a certain clinical case.

As we explained in the Introduction, the rationale of using semi-competing risk model is that in the prediction of AD/ADRD, the study population, which are usually the elderly population, is subject to both the risk of AD/ADRD and the risk of death because of age, so the competing risk of death would censor the AD/ADRD outcomes, thus serves as an informative censoring for AD/ADRD failure events. By using a semi-competing risk model, we hope to improve the performance/accuracy of predicting AD/ADRD. We have elaborated more on this in the Introduction section.

2. The second question is about the precision and sensitivity of the model on prediction of the risk of AD. Is this model reliable to predict AD risk? Please provide more details in the paper.

Thank you for your helpful comment on our study. We appreciate your question regarding the precision and sensitivity of our model for predicting the risk of AD. Our study utilized a large cohort of individuals with MCI and used a semi-competing risks regression model to predict the risk of AD/ADRD while accounting for the competing risk of death. The focus on this investigation is on the association parameter in the semi-competing risk model, which can help identify the subpopulation with high-risk of developing AD/ADRD. By allowing the latent frailty of the patients, we can properly account for the dependence between the progression from MCI to AD/ADRD, and the competing risk of death. As shown in survival analysis literature, this strategy can avoid the bias in the estimated coefficients, compared to the “standard Cox regression” where death is treated as a random censoring. [1,2]

We believe the proposed model, with appropriate handling of competing risk of death, is reliable in identifying/predicting the subpopulations with high risk or fast progression from MCI to AD/ADRD. On the other hand, our proposed model is not specified designed for patient-level risk prediction. Nevertheless, based on research from the semi-competing literature, it has been showed that semi-competing risk model demonstrated good discrimination and calibration in patient-level risk prediction. [3,4]

Reference

1           Fine JP, Jiang H, Chappell R. On semi-competing risks data. Biometrika 2001;88:907–19.

2           Haneuse S, Lee KH. Semi-competing risks data analysis: accounting for death as a competing risk when the outcome of interest is nonterminal. Circ Cardiovasc Qual Outcomes 2016;9:322–31.

3           Reeder HT, Lu J, Haneuse S. Penalized estimation of frailty‐based illness–death models for semi‐competing risks. Biometrics 2022.

4            Lee KH, Haneuse S, Schrag D, et al. Bayesian Semi-parametric Analysis of Semi-competing Risks Data: Investigating Hospital Readmission after a Pancreatic Cancer Diagnosis. J R Stat Soc Ser C Appl Stat 2015;64:253–73. doi:10.1111/RSSC.12078

Author Response File: Author Response.docx

Reviewer 3 Report

The authors did not address my major comment (the first comment), which is about the so-called "standard cox model" mistakenly described in this paper as a method that ignores the competing risks structure and has bias. In fact, it is known that if death is treated as censoring in the estimation, the obtained hazard ratio is for the cause-specific hazard, fully accounting for possible correlation between competing events. That is, the "standard cox model" is a fully valid method under the competing risk framework. I will not reiterate my original comment. Please read my original comment (the first comment) and address each of the points raised. Listing a few specific problematic statements related to this concern for the authors to consider:

 

- "it could lead to biased results and misleading conclusions if we simply treat death as censoring and fit a standard Cox proportional regression model."

 

- "and that the additional adjustment of the semi-competing regression model did not yield significant improvements. It is possible that death may not be a competing risk in the progression between MCI to AD/ADRD"

 

- "and the standard Cox model may not fully capture the association between these predictors and the risk of developing AD/ADRD 251 due to the presence of death as a competing risk."

Author Response

The authors did not address my major comment (the first comment), which is about the so-called "standard cox model" mistakenly described in this paper as a method that ignores the competing risks structure and has bias. In fact, it is known that if death is treated as censoring in the estimation, the obtained hazard ratio is for the cause-specific hazard, fully accounting for possible correlation between competing events. That is, the "standard cox model" is a fully valid method under the competing risk framework. I will not reiterate my original comment. Please read my original comment (the first comment) and address each of the points raised. Listing a few specific problematic statements related to this concern for the authors to consider:

- "it could lead to biased results and misleading conclusions if we simply treat death as censoring and fit a standard Cox proportional regression model."

- "and that the additional adjustment of the semi-competing regression model did not yield significant improvements. It is possible that death may not be a competing risk in the progression between MCI to AD/ADRD"

- "and the standard Cox model may not fully capture the association between these predictors and the risk of developing AD/ADRD due to the presence of death as a competing risk."

Thanks for your helpful comments. However, we do not entirely agree with the statement that “the "standard Cox model" is a fully valid method under the competing risk framework”. While the standard Cox model can be used to estimate the effect of risk factors on the cause-specific hazard, it does not fully account for the possible correlation between competing events. In the presence of competing risks, treating death as random censoring can lead to biased estimates of the cause-specific hazard and hazard ratios. Such limitation of standard Cox model has been extensively discussed in the literature; see, for example, Fine et al., (2001), Haneuse et al., (2016), and others. [1,2,5]This is because the standard Cox model assumes that censoring is non-informative, meaning that the probability of being censored is not related to the risk of experiencing the event of interest or the competing event. However, in the presence of competing risks, this assumption may not hold, and the probability of being censored may be related to the risk of experiencing the competing event.  We have described our rationale of using semi-competing risk approach extensively in the Introduction section and throughout the manuscript. Our model output suggest that death could be non-informative and may not bias the prediction of AD/ADRD, but that was not the underlying research question that we aimed to answer.

Reference

1           Fine JP, Jiang H, Chappell R. On semi-competing risks data. Biometrika 2001;88:907–19.

2           Haneuse S, Lee KH. Semi-competing risks data analysis: accounting for death as a competing risk when the outcome of interest is nonterminal. Circ Cardiovasc Qual Outcomes 2016;9:322–31.

3           Reeder HT, Lu J, Haneuse S. Penalized estimation of frailty‐based illness–death models for semi‐competing risks. Biometrics 2022.

4           Lee KH, Haneuse S, Schrag D, et al. Bayesian Semi-parametric Analysis of Semi-competing Risks Data: Investigating Hospital Readmission after a Pancreatic Cancer Diagnosis. J R Stat Soc Ser C Appl Stat 2015;64:253–73. doi:10.1111/RSSC.12078

5           Peng M, Xiang L. Correlation-based joint feature screening for semi-competing risks outcomes with application to breast cancer data. doi:10.1177/09622802211037071

 

Author Response File: Author Response.docx