OHDSI Home | Forums | Wiki | Github

Phenotype Phebruary Day 5- Alzheimer's Disease


Day 5 of Phenotype Phebruary. Still lots of methodological topics to discuss and disease areas to investigate. Today, I’ll try to start a conversation of the phenotype that was most highly voted on across our community: Alzheimer’s disease.

Clinical description:

Alzheimer’s disease is a progressive neurodegenerative disorder and the most common cause of dementia (loss of cognitive functions interfering with daily activities), representing 60-80% of cases (according to Alzheimer’s Association). Intitial symptoms of Alzheimer’s disease may be short-term memory loss and other difficulties associated with mild cognitive impairment, such as word-finding, visual/spatial issues, and general confusion. Diagnosis of Alzheimer’s disease may involve neurological exam, including brain MRI or CT scans, to identify other potential causes of dementia other than Alzheimer’s, and mental cognitive status tests. Drugs approved for use in Alzheimer’s disease include cholinesterase inhibitors (such as donepezil, galantamine, or rivastigmine) and memantine, which are primarily aimed at treating cognitive symptoms. In 2021, aducanumab was approved by US FDA on the basis of clinical trial data suggesting reduction of amyloid beta plaque. Alzheimer’s disease risk increases with age, with most cases detected after 65 years old. Prevalence of AD is higher in females than males, though that is attenuated by female longer life span. It is one of leading causes of death globally, and second-leading cause in high-income countries (WHO).

Phenotype development:

I’ve mentioned in prior posts that a valuable starting point for phenotype development can be the published literature, and I’ve shown how you - provided that a journal article supplied enough details - you can replicate their algorithms using OHDSI tools. But I want to take a digression here for a little rant: if observational researchers all need to develop phenotypes to conduct our analyses and should all review prior literature as part of our research process, then why is so hard to search for publications of observational research and extract out the phenotypes that were previously used? If phenotypes are so central to the integrity of our research, then as a research community, why do we accept short freetext descriptions of phenotypes in manuscripts, sometimes without list of codes and often without a complete specification of the logic that was used to implement them? And for those of us promoting increased transparency, when we try to add additional detail in supplemental materials, why do we often format it in ways that make it painful for others to re-use without extensive manual curation? When I read a paper that I’m excited by and want to replicate, what I wouldn’t give to just get a JSON specification of conceptsets and cohort logic, not only to save me effort but also to prevent error and avoid ambiguity.

The first step is actually finding papers that contain phenotypes to consider. Now, PubMed is an AMAZING resource; my wife is an English professor, so I get to see what research can require if one doesn’t have a centralized repository of published scholarship (it looks aweful), and I never take for granted the ability to quickly identify and retrieve scientific references. Thank you NLM! But, perhaps because we are in a fairly niche space, the maturity of observational database research publications and their content isn’t always straightforward to capture.

The search problem here is: Find all observational database studies that contain a phenotype algorithm for Alzheimer’s disease. A plea to our community: if anyone has a great PubMed search strategy for this task, please share it by replying to this post.

I’ll share the search strategy I use, which is something that I started working on back in the OMOP days and cleaned up and improved with the help of Gayle Murray, our wonderful library expert at Janssen.

(“Alzheimer disease”[MeSH Terms] OR “Alzheimer”[Title/Abstract])
((“retrospective cohort”) OR (Epidemiology[MeSH Terms]) OR (Epidemiologic Methods[MeSH Terms]) OR (phenotype[Title/Abstract]) OR (insurance) OR (claims) OR (database) OR (Diseases Category/epidemiology[MeSH Terms]) OR (Validation Study[Publication Type]) OR (Validation Studies as Topic[MeSH Terms]) OR (Sensitivity and Specificity[MeSH Terms]) OR (Predictive Value of Tests[MeSH Terms]) OR (Reproducibility of Results[MeSH Terms]) )
((Medicaid) OR (Medicare) OR (Truven) OR (Optum) OR (Medstat) OR (“Nationwide Inpatient Sample”) OR (“National Inpatient Sample”) OR (PharMetrics) OR (PHARMO) OR (ICD-9[Title/Abstract]) OR (ICD-10[Title/Abstract]) OR (IMS[Title/Abstract]) OR (“electronic medical records”[Text Word]) OR (Denmark/epidemiology[MeSH Terms]) OR (Veterans Affairs[Title/Abstract]) OR (“Premier database”[Title/Abstract]) OR (“National Health Insurance Research Database”[Title/Abstract]) OR (Outcome Assessment[Title/Abstract]) OR (“insurance database”[Title/Abstract]) OR (Database Management System[MeSH Terms]) OR (Medical Records Systems, Computerized[MeSH Terms]) OR (“Positive predictive value”[Title/Abstract]) )
(“Clinical Trial”[pt] OR “Editorial”[pt] OR “Letter”[pt] OR “Randomized Controlled Trial”[pt] OR “Clinical Trial, Phase I”[pt] OR “Clinical Trial, Phase II”[pt] OR “Clinical Trial, Phase III”[pt] OR “Clinical Trial, Phase IV”[pt] OR “Comment”[pt] OR “Controlled Clinical Trial”[pt] OR “Letter”[pt] OR “Case Reports”[pt] OR “Clinical Trials as Topic”[Mesh] OR “double-blind”[All] OR “placebo-controlled”[All] OR “pilot study”[All] OR “pilot projects”[Mesh] OR “Prospective Studies”[Mesh] OR “Genetics”[Mesh] OR (“Genotype”[Mesh]) OR (biomarker[Title/Abstract]))

Basically, the search string is a combination of 1. the phenotype target (varients in MeSH terms or freetext), 2. a list of markers to indicate the type of study, 3. a list of terms to verify that its a database study, and 4. exclusions to remove non-observational research.

This returns back 1,009 results, most of which are not helpful for that I’m trying to accomplish. So, I’ve got a sensitivity/specificity problem with my search string…just like my phenotypes :slight_smile:

The second step is to extract phenotype algorithms from the relevant papers. Here are a few papers that were near the top of the list or were , which are useful for this discussion:

McCarthy et al, " Validation of Claims Algorithms to Identify Alzheimer’s Disease and Related Dementias", J Gerontol A Biol Sci Med Dec2021.

They provide a brief description of their three algorithms with the following table to enumerate code choices (shown below) . In their decription, they state that they are starting with a codeset from prior literature, but then they evaluate modifications by “adding diagnosis codes for dementia with Lewy Bodies (331.82), other cerebral degeneration (331.89), and other nonspecified senile psychosis (290.8) based on discussion with experts in the field”

Jain et al., " Using Medicare claims in identifying Alzheimer’s disease and related dementias", Alzheimers Dement Oct2020.

The authors describe that “To ensure the correct list of codes, we began with a list of ADRD
diagnostic codes used in the literature”, but then, “We also considered codes for other medical
conditions, such as other degenerative conditions, delirium”.

Kern et al, “Treatment with TNF-α inhibitors versus methotrexate and the association with dementia and Alzheimer’s disease” Alzheimers Dement Sept2021.

The authors describe " AD tends to be under‐coded in claims data and thus we used the broader definition of “dementia” as our primary endpoint, which includes AD as well as generic conditions of “senility,” “degenerative brain disorder,” and others, but excludes Lewy body dementia, drug‐ and alcohol‐induced dementia, or dementia caused by concussions or syphilis, among others. A full list of the included concepts and mapped ICD‐9‐CM and ICD‐10‐CM codes can be found in Appendix Table.", and then provide this:

Code Name Vocabulary
Alzheimer’s Disease codes
331.0 Alzheimer’s disease ICD9CM
G30.0 Alzheimer’s disease with early onset ICD10CM
G30.1 Alzheimer’s disease with late onset ICD10CM
G30.8 Other Alzheimer’s disease ICD10CM
G30.9 Alzheimer’s disease, unspecified ICD10CM
Dementia Codes
290.0 Senile dementia, uncomplicated ICD9CM
290.10 Presenile dementia, uncomplicated ICD9CM
290.11 Presenile dementia with delirium ICD9CM
290.12 Presenile dementia with delusional features ICD9CM
290.13 Presenile dementia with depressive features ICD9CM
290.20 Senile dementia with delusional features ICD9CM
290.21 Senile dementia with depressive features ICD9CM
290.3 Senile dementia with delirium ICD9CM
290.40 Vascular dementia, uncomplicated ICD9CM
290.41 Vascular dementia, with delirium ICD9CM
290.42 Vascular dementia, with delusions ICD9CM
290.43 Vascular dementia, with depressed mood ICD9CM
294.0 Amnestic disorder in conditions classified elsewhere ICD9CM
294.10 Dementia in conditions classified elsewhere without behavioral disturbance ICD9CM
294.11 Dementia in conditions classified elsewhere with behavioral disturbance ICD9CM
294.20 Dementia, unspecified, without behavioral disturbance ICD9CM
294.21 Dementia, unspecified, with behavioral disturbance ICD9CM
331.0 Alzheimer’s disease ICD9CM
331.11 Pick’s disease ICD9CM
331.19 Other frontotemporal dementia ICD9CM
331.2 Senile degeneration of brain ICD9CM
331.9 Cerebral degeneration, unspecified ICD9CM
797 Senility without mention of psychosis ICD9CM
F01.50 Vascular dementia without behavioral disturbance ICD10CM
F01.51 Vascular dementia with behavioral disturbance ICD10CM
F02.80 Dementia in other diseases classified elsewhere without behavioral disturbance ICD10CM
F02.81 Dementia in other diseases classified elsewhere with behavioral disturbance ICD10CM
F03.90 Unspecified dementia without behavioral disturbance ICD10CM
F03.91 Unspecified dementia with behavioral disturbance ICD10CM
G30.0 Alzheimer’s disease with early onset ICD10CM
G30.1 Alzheimer’s disease with late onset ICD10CM
G30.8 Other Alzheimer’s disease ICD10CM
G30.9 Alzheimer’s disease, unspecified ICD10CM
G31.01 Pick’s disease ICD10CM
G31.09 Other frontotemporal dementia ICD10CM
G31.1 Senile degeneration of brain, not elsewhere classified ICD10CM
G31.89 Other specified degenerative diseases of nervous system ICD10CM
G91.4 Hydrocephalus in diseases classified elsewhere ICD10CM

[Related aside, but this is a very nice paper from collaborators within our OHDSI community, @Dave_Kern and @scepeda. It shows how OHDSI tools and best practices can be used to conduct a population-level effect estimation study across two claims databases to test a biological hypothesis. Nice examples of large-scale propensity score adjustment to achieve covariate balance and use of negative controls to evaluate residual confounding]

So, three papers trying to use claims data to study Alzheimer’s disease and related dementias, three different formats for how the codelists are shared, and indeed, three different lists of codes. But how does one go about finding the subtle differences between the codelists? And how does one determine if the differences are actually impactful?

One approach that I’ve created as a bit of a habit for myself, when reviewing the prior literature and trying to create a conceptset: I think about the literature as the union of all codes that I may be interested in, and I make a conceptset that maps from the source codes provided to their corresponding OHDSI vocabulary standard concepts, so that ultimately I end up with a conceptset expression that subsumes all the codes from the literature (and by virtue of the SNOMED hierarchy and source code mappings, may also sweep in a few extra codes that the literature had missed).

What does my standardized conceptset expression look like here, to capture all the ideas from McCarthy, Jain and Kern?

If we look at the ICD9CM codelist covered, we can see 40 codes provided:

And for ICD10CM, we’ve got 26 codes:

But, you don’t want these teeny tiny screenshots any more than I want to have to manually type out code-by-code from the various forms in the publication. What you want is something that is both human-readable and computer-executable. Here’s the JSON specification for the dementia conceptset (xml filetype only to enable it to be uploaded here on the forums)
Phenotype Phebruary dementia conceptset.xml (9.6 KB)

Once I have created my conceptset covering the literature, then I can continue with some of the other steps that we’ve discussed earlier this week, like using PHOEBE to see if any other concepts are recommended, creating cohorts using those conceptsets, running CohortDiagnostics or PheValuator to evaluate the cohorts, etc.

Now, the ‘dementia’ conceptset may look not-so-straightforward, and that’s in part because the various papers disagreed on assorted edge cases of other types of dementias and cognitive decline. But the conceptset for ‘Alzheimer’s disease’ is a thing of beauty:

One concept, plus its descendants, covers 34 standard concepts and fully subsumes all the ICD9CM and ICD10CM codes highlighted by the papers.

I want to go back to the paper by @Dave_Kern and @scepeda . They do something quite smart in their analysis: they are interested in the target of Alzheimer’s disease, but they noted from prior literature that AD algorithms had low sensitivity but good specificity. But we also know from prior literature that Alzheimer’s makes up 60-80% of dementia cases. So, one can consider a dementia phenotype to be a ‘more sensitive, less specific’ phenotype algorithm for Alzheimers. Given that these two alternative algorithms represent a logical tradeoff in measurements errors, they executed their analyses using both definitions. And when they generated consistent estimates with both approaches, they could more confidently conclude that their finding wasn’t just a bias due to one algorithm. In situations where there is no one clear right algorithm, and the impact of alternative definitions can be substantial (as was here), then this idea of ‘bounding’ your analyses with a sensitive algorithm and a specific algorithm seems like a good strategy to consider when trying to manage worries of measurement error.

Resources built today to help this discussion (conceptsets are embedded within these cohorts):

I know several of you have done research in Alzheimers and dementia. Share your phenotype experience. What did you learn? How can future work be improved? Is there anything that I’ve said here that resonated or that you disagree with? Let’s discuss!

Here are the PheValuator results for the two Alzheimers disease algorithms:

We see that the less specific definition, including all of dementia and its descendants, had lower PPV in both datasets examined, especially pronounced in Medicaid (low socioeconomic status individuals). The loss in PPV in Medicaid was accompanied by a jump in sensitivity so there is a trade-off there. In Medicare, generally those > 65YO, there was little change in sensitivity.

This is really cool @jswerdel ! Fascinating to see an example of two databases side-by-side where in one, there is little change in sensitivity and only a modest decrease in PPV, but in the other, a sizable shift in both operating characteristics. This underscores the value of systematically applying a tool like PheValuator across multiple definitions and multiple databases, otherwise I don’t know how you would find an insight like this.

Great work everyone - we are actually funded by the NIA to do some chart review based validation on dementia algorithms, so will be using this one as an example.
In addition, I’ll follow-up with Joel about PheValuator to understand it - the sens and PPV is much higher than one would usually get with a reference standard.
I wonder if you all would be willing to take a survey if you have used a dementia algorithm already - details are below. We’d welcome your input!

This survey is being conducted as part of the activities of the NIA IMPACT Collaboratory Technical Data Core. It is designed to gather information about the current state of using algorithms to identify patients with dementia in clinical populations. We are surveying a variety of institutions and asking about your local experiences and expertise in this domain. This survey is only about your professional knowledge and was determined to be non-human subjects research by the OHSU IRB.
For the purposes of this survey, “you” should be interpreted as someone experienced with pulling data for research studies at your institution: it is generally the informatics or IT lead who manages the research or clinical data warehouses. If you feel like you are not the best person in your institution to answer these questions, please feel free to forward to the person who is more appropriate.
It is also okay to have multiple people from your institution contribute to filling out the survey, if needed. Collaborations can be combined into a single survey response or each person can fill out their own version, addressing the questions that they are knowledgeable about answering.


Thank you!