Clinical factors associated with progression to dementia in people with late-life depression: a cohort study of patients in secondary care

Objectives Depression can be a prodromal feature or a risk factor for dementia. We aimed to investigate which clinical factors in patients with late-life depression are associated with a higher risk of developing dementia and a more rapid conversion. Design Retrospective cohort study. Setting South London and Maudsley NHS Foundation Trust (SLaM) secondary mental healthcare services. Participants The SLaM Clinical Record Interactive Search was used to retrieve anonymised data on 3659 patients aged 65 years or older who had received a diagnosis of depression in mental health services and had been followed up for at least 3 months. Outcome measures Predictors of development of incident dementia were investigated, including demographic factors, health status rated on the Health of the National Outcome scale for older people (HoNOS65+), depression recurrence and treatments including psychotropic drugs and cognitive behavioural therapy (CBT). Results In total, 806 (22.0%) patients developed dementia over a mean follow-up time of 2.7 years. Significant predictors of receiving a dementia diagnosis in fully adjusted models and after accounting for multiple comparisons were older age (adjusted HR=1.04, 95% CI 1.03 to 1.06 per year difference from sample mean) and the HoNOS65+ subscale measuring cognitive problems (HR=4.72, 95% CI 3.67 to 6.06 for scores in the problematic range). Recurrent depressive disorder or past depression (HR=0.65, 95% CI 0.55 to 0.77) and the receipt of CBT (HR=0.73 95% CI 0.61 to 0.87) were associated with a lower dementia risk. Over time, hazards related to age increased and hazards related to cognitive problems decreased. Conclusions In older adults with depression, a higher risk of being subsequently diagnosed with dementia was predicted by higher age, new onset depression, severity of cognitive symptoms and not receiving CBT. Further exploration is needed to determine whether the latter risk factors are responsive to interventions.

Page 11, lines 23-25. Am I correct here that the people in the analysis are solely the 806 with both a depression and dementia diagnosis as seems to be implied here? If so I can't see that since all the people in the analysis have an 'event' occurring namely dementia why you didn't just perform a linear regression with time from depression to dementia as the outcome variable since everyone in the analysis has experienced a diagnosis of dementia and thus has a time when dementia took place so you have no need for a Cox Model since you have no censored observations (ie there are no people in the analysis who did not have a diagnosis of dementia in the time frame of the study). In fact a better approach would be to additionally take advantage of the extra information in censored observations (people with depression but who have not as yet being diagnosed with dementia but are assumed to definitely get dementia at some point) and use the Cox Model on all 5910 patients diagnosed with depression including the 5104 (Page 11, lines 14-19) who had received a diagnosis of depression but had not as yet received a diagnosis of dementia. You would put time from depression diagnosis as the outcome with a binary censored indicator (either no dementia diagnosed by that time or dementia was diagnosed at that time for each patient). I would add the sample size (N=806 or N=5910 or other) in Table 2 (Page 15) as is done in Table 1 (Page 12) to clarify who are the people in the Cox Model analysis.
Page 15. The results for the three models look very similar. I would therefore just fit Model 3 with contains all the covariates and which appears to address the associations of interest as reported in the abstract on page 3. I assume you have tested the proportional hazards assumption underlying the use of covariates with the Cox Proportional Hazards Model? Namely that the hazard rate over time is the same for different covariate subgroups e.g. males and females.
I assume that the covariate values do not change over time e.g. married or cohabiting status has not changed over the time period. If there are covariates changing over the time period would you consider fitting them as time varying covariates? I also wondered, on a similar theme if you could have alternatively modelled the depression covariate as a time-varying covariate since people who have been diagnosed as depressed would have come out of depression at a later date hence their depression status would have changed over time.

REVIEWER
Lynette Chenoweth University of New South Wales, Australia REVIEW RETURNED 20-Dec-2019

GENERAL COMMENTS
The study objectives and procedures are clearly described, the statistical methods are suitable and the results answer the research questions. The manuscript is well written and logically structured. The Discussion of study findings is particularly thoughtful and insightful. I cannot recommend any improvements that are required.

REVIEWER
Prof Stephen Z Levine (PhD) University of Haifa REVIEW RETURNED 23-Jan-2020

GENERAL COMMENTS
This manuscript examines factors associated with dementia among persons with depression. Despite my remarks, I think this is an interesting contribution to this domain of research.
1. The abstract states-"Depression in later life is increasingly being recognized as a prodrome for the development of dementia" -this should read "There is evidence that depression is a prodrome factor for the risk of dementia" -the jury is not out and it may be a risk factor. 2. The manuscript might benefit by mentioning the projected increase in depression.1 3. In the context of depression studies I do not think 2.8y is "relatively long" compared with high-quality epidemiology studies (not clinical trials). 4. In today's age of big data I would not say this is the accurate-"unique access to anonymized clinical data" remove "unique". 5. The authors should re-evaluate the evidence regarding antidepressants. They cite the Lancet 2017 paper, yet ignored the WHO guidelines and large observational studies. First, recent large observational studies find increased risk (uncited so cite2). Many thanks for those interesting observations and suggestions, which have led to us reconsider our approach. As you highlighted, we didn't want to lose all the observation from the censored observation. As censoring point, we chose the last clinical contact with mental health services, as this is a point when we can be reasonably certain that the person has not (yet) developed dementia. In line with the dementia ascertainment, we only included patients who were followed-up for at least 3 months, which changed the numbers slightly. Missingness was higher in those not developing dementia, leading to the necessity to impute missing values.
This affects all chapters of the manuscript, and we have highlighted these in yellow. Our three models remain similar, but we still think there is some value for the reader to see all three models; especially as following the suggestions from Reviewer 2, we have clarified whether a significance value would hold following correction for multiple comparisons.

Reviewer 2:
The study objectives and procedures are clearly described, the statistical methods are suitable, and the results answer the research questions. The manuscript is well written and logically structured. The Discussion of study findings is particularly thoughtful and insightful. I cannot recommend any improvements that are required.
Many thanks for the positive feedback and very encouraging comments.
The abstract states-"Depression in later life is increasingly being recognized as a prodrome for the development of dementia" -this should read "There is evidence that depression is a prodrome factor for the risk of dementia"the jury is not out and it may be a risk factor.
Many thanks, we updated the abstract to reflect this uncertainty.

2.
The manuscript might benefit by mentioning the projected increase in depression.1 Many thanks for highlighting this, this certainly adds to the importance of our study and we have added this to the introduction.

3.
In the context of depression studies I do not think 2.8y is "relatively long" compared with high-quality epidemiology studies (not clinical trials).
Many thanks, we have removed this point from the article summary.
In today's age of big data I would not say this is the accurate-"unique access to anonymized clinical data" remove "unique". Many thanks for pointing us to these important resources, which has facilitated a rewrite of the respective section of the discussion.

6.
Clarify this is actually incident dementia.
We added this to the abstract and the methods.
Abstract, Methods

7.
How were the competing risks of dementia and mortality accounted for.
We used Cox regression models to account for censoring due mortality or end of follow-up time. We changed our approach in response to suggestion 2 by Reviewer 2, and clarified in the methods: Date of first depression diagnosis after the age of 65 years served as the index date. Patients were followed up until the date of the first dementia diagnosis (incident dementia) given in the health record or until the last face-to-face contact with a professional from secondary mental health services (clinical event or appointment). Further, follow-up also ended at death or a censoring point on 30/06/2017. Patients were excluded if they had a recorded dementia diagnosis before or within 3 months after index date or if they had less than 3 months of contact with SLaM services after a depression diagnosis.

8.
How were violations of the assumption of proportional hazards accounted for? Maybe a Zou Poisson model would be superior here?
We prefer the Cox proportional hazard model, as not all covariates are binary, and the Cox model has less assumptions. We tested for violations of the proportional hazard Methods assumptions and added variables as time-varying if they were violated.

11.
Are these prescribed / prescribed & purchased / reported antidepressant use? Each method with their own restrictions.
Medication prescription is ascertained via natural language processing from free text. Together with being unable to account for the time-dependent nature of prescribing, we acknowledged the drawbacks of this approach as follows in the limitations: Medication prescription was ascertained either within the 2 years before or after the index date of first depression diagnosis through natural language processing, but the timedependent nature of prescribing could not be accounted for. Further, the output of natural language processing depends on the accuracy and quality of data entry, which varies by individual clinician and is compromised through the use of jargon, idiosyncratic abbreviations or misspellings 17 . Although it has been shown that precision and recall are relatively high for the medication natural language processing application 17 , there remains a risk under-or overestimating the true prescribing prevalence, especially as adherence to medication cannot be established 54 .
The way the medications are analyzed does not address their time-depend nature.

10.
Age -often studies add a linear and quadratic terms to account for the change in risk with age.
Many thanks, this is a great suggestion. We conducted a sensitivity analysis and added the following to the Results:

12.
Use FDR P-values owing to multiple tests.
Also, many thanks for making this important point. For Table 2 we now report p<.05 and a Bonferroni corrected p-value of p<.002.
Models of increasing complexity -Good approach -As we are assessing various predictors, we'd like to report all please include the BICs and other information fit indices & just report the most parsimonious model. models. We have however used the robust estimate of variance to account for the same subjects appear repeatedly in the risk pools.

14.
Please expand on this -353 (43.8%) developed Alzheimer's disease. If we believe the literature AD accounts for far more cases of dementia than this figure.
This is because we only look at first recorded dementia subtype; to explain this we added the following to the discussion: Of the 806 patients diagnosed with dementia, we found that 25% received an initial diagnosis of unspecified dementia and only 44% a diagnosis of Alzheimer's disease. This is lower than the expected incidence of Alzheimer's disease in two-thirds of dementia cases 51 and likely to reflect the naturalistic nature of the data, whereby the specific subtype is often recorded as unspecified until further investigations are conducted or more specific symptoms emerge. Subsequently the recording of specific subtypes increases if the whole record is considered 52 53 .

15.
Confounders are inadequately accounted for and the authors correctly acknowledge this shortcoming.
Many thanks, we have now made this clearer in the limitations, in addition to the Highlights section.
Article summary,

16.
Considering the suicide rate of old people with depression, is selection mortality possible?
That is an important point. In initial Cox analyses we consider mortality as censoring point, but not in the sub-cohort of 806 patients who developed dementia. We have acknowledged this in the limitations as follows: Additionally, late-life depression confers an increased mortality risk, both in relation to suicide and when suicidality is accounted for 50 , and the sub-cohort of patients who received both depression and dementia diagnoses in SLaM was selective to the 806 patients who survived long enough to receive a dementia diagnosis.  Table 2   Figure 2   Supplementary Table 1 Page 10, lines 46-50. Does it, therefore, follow from the creation of 19 imputed data sets that the results in Tables 1 (descriptives), 2 (Cox model) and 3 (cognitive score mean times to dementia) are based upon pooled estimates averaged across the multiple Yes, that is correct. We have now clarified in the methods that we used the STATA mi package and combined coefficients using Rubin's rules.

Methods -Statistical Analysis
imputations? If so, what procedure did you use in STATA to do the pooling across imputations? A brief mention of this in the paper would help understanding of what you have done. I have not come across multiple imputation using the Cox model before and am not familiar with the multiple imputation references given in the paper but I would imagine you can pool estimates and their covariance matrices across the imputations as you would do for any other regression model.
Page 14, lines 11-12 and 27-38 and Page 15, Table 2 and line 47. Slight confusion here in that you mention that you are only considering p-values less than 0.002 as significant to account for multiple comparisons (Page 14, lines 11-12) yet you go on to flag two variables, having hallucinations and taking antidepressants which both have p-values above 0.002 as associated with lower dementia and then say (Page 20, lines 50-53) that using multiple comparison corrections are conservative.
Many thanks, this is an important observation. Although having hallucinations and taking antidepressants were not significant predictors after accounting for multiple comparisons, we felt that they were nonetheless worth some discussion. To avoid confusion, we have removed the sentence referring to FDR procedures as being conservative in survival analyses.

Discussion
final paragraph Page 15, Table 2 line 14. I wondered if you had centred age (ie subtracted its mean) to make its incremental effect as measured by the Cox Model more interpretable? If you do this then you are automatically comparing changes of one year increases in the hazard rate to the more intuitive average age rather than age zero (birth).
We hadn't done this as yet, but it is a good idea. It is surprising that this is not done more often, and so we have added an explanation to the manuscript.

Methods -Covariates
Results Table 2 Page 30, Figure 2. I don't see mention in the figure of any footnotes explaining to what the 'a' and 'd' superscripts alongside some of the variables mean.
Apologies, the footnotes/legends for the figures can be found at the end of the main text document and had been entered in the submission portal, but these are not displayed with the figures themselves at this stage. We have been advised by the editorial office that this is the correct approach for submission.