Alzheimer’s Testing and Diagnosis
Alzheimer’s testing will be modeled as an intervention with no affected outcomes which introduces PET, CSF, and BBBM testing for eligible predementia populations, as defined in Reference Scenario and Alternative Scenario 1.
Abbreviation |
Definition |
Note |
|---|---|---|
BBBM |
Blood-based biomarker |
BBBM tests measure blood plasma protein levels and are less invasive than CSF or PET tests |
CSF |
Cerebrospinal Fluid |
CSF tests require a lumbar puncture |
PET |
Positron Emission Topography |
PET scanners measure brain protein levels are only available at specialized medical facilities |
Intervention Overview
Alzheimer’s testing is classified into two groups - existing (CSF and PET) testing, and a hypothetical BBBM testing. Existing testing already exists in present-day conditions, and our simulation will investigate the potential impact of the introduction of a novel BBBM test in Alternative Scenario 1. The introduction of BBBM testing will replace some of the existing tests in Alternative Scenario 1, and will also be used to inform treatment in Alternative Scenario 2.
Vivarium Modeling Strategy
Note that disease model will take timestep before testing model, so simulants can be tested on incidence to new stage.
Existing testing - CSF and PET
CSF and PET are existing biomarker tests which can identify cases of Alzheimer’s disease. We model CSF and PET tests in our Reference Scenario in order to estimate the number of these more expensive, invasive tests which would be replaced by the introduction of BBBM tests in Alternative Scenario 1.
Location-specific testing rates
CSF and PET testing rates in the US, Germany, Spain, the UK and China are given by [Roth-et-al-2023-Diagnostic-Pathways]. The paper surveys health care providers about patients presenting with cognitive complains. Patients with cognitive complaints presenting to a health care provider is the denominator for these test rates.
Swedish CSF testing rate is from [Falahati-et-al-2015-SveDem], and PET comes from that value and the relative CSF vs PET test proportion (90% CSF, 10% PET) given by [Mattke-et-al-2024-Sweden-Capacity]. The cohorts for these studies are incident AD dementia cases and patients with confirmed MCI respectively.
For Japan, Israel, Taiwan and Brazil, because we have not yet found CSF and PET testing rates for these locations, we use the total CSF and PET testing rates across all countries from [Roth-et-al-2023-Diagnostic-Pathways].
After choosing these mean values, we subtract 50% for a lower confidence bound and add 50% for an upper confidence bound to reflect substantial uncertainty (to be used in parameter uncertainty draws).
Note that the three sources for these test rates have slightly different cohorts and therefore test rate denominators. These test rates will be applied to simulants with MCI or AD dementia.
Location |
CSF mean % (Confidence) |
PET mean % (Confidence) |
|---|---|---|
US |
10.8 (5.4, 16.1) |
15.0 (7.5, 22.5) |
Germany |
18.7 (9.4, 28.1) |
16.0 (8.0, 24.0) |
Spain |
24.6 (12.3, 36.9) |
25.9 (12.9, 38.8) |
Sweden |
40.5 (20.3, 60.8) |
4.5 (2.3, 6.8) |
UK |
9.5 (4.7, 14.2) |
10.7 (5.3, 16) |
Japan |
13.3 (6.7, 20) |
14.9 (7.5, 22.4) |
Israel |
13.3 (6.7, 20) |
14.9 (7.5, 22.4) |
Taiwan |
13.3 (6.7, 20) |
14.9 (7.5, 22.4) |
Brazil |
13.3 (6.7, 20) |
14.9 (7.5, 22.4) |
China |
4.4 (2.2, 6.6) |
6.1 (3, 9.1) |
Implementation
Each simulant will be assigned a propensity for receiving a test (0 to 1). A low propensity value means the simulant is likely to receive a test, while a high propensity value means the simulant is unlikely to receive a test. The propensity will apply for the simulant’s lifetime.
On timestep
On each timestep, use the following steps to assign CSF and PET tests:
Assess eligibility based on the following requirements:
Simulant is in MCI stage or AD dementia stage (though due to lack of age requirement, no simulants should be tested in AD dementia stage - only on MCI incidence)
Simulant has never received a CSF or PET test before
Simulant has never received a positive BBBM test before
If eligible (meets all requirements), check propensity. If the propensity value is less than the location-specific testing rate (CSF rate plus PET rate), give the test. If not, do not give the test. Do not assign a diagnosis.
If the simulant receives a test, assign whether it is a CSF or PET test based on location-specific rates, independently of the testing propensity and other random choices. More explicitly, given that a simulant receives a test, the probability of getting a CSF test should be (CSF rate) / (CSF rate + PET rate), and the probability of getting a PET test should be (PET rate) / (CSF rate + PET rate).
On initialization
To avoid large numbers of simulants being tested on the first simulation time step, we must initialize simulant test history status so that some number of simulants have already been tested at simulation start. Only simulants who were not eligible for testing at simulation start, but become eligible after the first time step, should be tested at the first time step.
To accomplish this, simulant eligibility should be checked at simulation initialization, and simulants who satisfy all eligibility requirements at that time should be marked as having previously received a CSF/PET test. These simulants will be ineligible for future CSF/PET testing.
Assumptions and Limitations
A simulant with an eligible propensity will be tested at the first time step they satisfy the stage and age criteria, and then can never be tested again, so propensity does not need to be re-assigned at any point;
Assume no testing in pre-clinical state;
Not used to assign treatment (no diagnosis);
Eligibility requirements impact the number of tests. The earlier the stage simulants are tested in, the more tests will be conducted (eg mild stage compared to MCI). The wider the age range, the more tests will be conducted (eg no age requirements vs 60-80 year olds);
Assumes no one gets both a CSF and PET test;
The testing happens quite rapidly, during the first six months of MCI; a future enhancement to this model might include a random chance of testing among those with propensity for getting tested, so that some fraction of testing happens later in the progression of the disease.
BBBM testing
BBBM testing is a hypothetical biomarker test which we will model in Alternative Scenario 1. It will replace some CSF/PET testing and assign positive/negative diagnosis which will inform treatment in Alternative Scenario 2.
Time-specific testing rates
Testing rates do not vary by location, age or sex. In 2020, 0% of eligible simulants are tested annually. This becomes nonzero in 2027, increasing to 10% at year 2030, then increases linearly for awhile, then levels off and eventually maxes out at 60% after 2045. We will model this as a piecewise linear function with knots at the following (year, coverage) values:
(2020.0, 0%)
(2027.0, 0%) – Note that this is 0% at the beginning of 2027, but coverage will become positive on the second time step that year
(2030.5, 10%) – Note that this is 10% at mid-year
(2045.0, 50%)
(2055.0, 60%)
(2100.0, 60%)
Eligibility for BBBM testing
A simulant is eligible for a BBBM test if they meet the following requirements:
Simulant is not in MCI or AD dementia state (they can only be in susceptible or preclinical)
Simulant age is \(\ge 65\) and \(< 80\)
Simulant has not received a BBBM test in the last three years (more precisely, they have not had a BBBM test on any of the previous five time steps)
Simulant has never received a positive BBBM test
Implementation
The simulant’s existing CSF/PET testing propensity will also be used as their BBBM testing propensity. This will cause CSF and PET testing to be displaced as BBBM testing scales up. At the client’s request, we will retest simulants every 3-5 years, rather than having all simulants be retested at a fixed interval of 3 years (which can cause unrealistic oscillations in the number of tests over time). In the implementation below, we assume that the time between tests is uniformly distributed in the interval \([3, 5]\) years.
On initialization
In order to avoid having an unreasonably large fraction of eligible simulants be tested immediately upon entering the simulation, we will assign a future BBBM test date to each initialized simulant who otherwise would have an opportunity for BBBM testing on their first time step.
This future BBBM test date assignment should meet the following requirements:
A next test date is assigned to simulants who meet the eligibility requirements for BBBM testing and have a testing propensity is less than the current BBBM testing rate. For simulants who don’t meet both these requirements, assign “not a time” (NaT) for their next test date.
The future BBBM test date should mirror the testing scheme of uniformly random retesting every 3-5 years.
Note
Implementation: We achieve the second criterion above by having two random draws. First, a uniformly random time \(W\) between 3 and 5 years is selected. This value is the time the simulant was assigned to wait to retest. Second, a uniformly random time \(T\) between zero and \(W\) is picked. This is the amount of time the simulant has been waiting so far. The time until the next test date is then calculated from these two draws as \(W - T\). For example, if in the first draw we select 4 years, and in the second draw we select 1.5 years, the simulant would be scheduled to retest at \(4-1.5 = 2.5\) years after entering the simulation.
Note: Although the simulation does not explicitly track a “prior BBBM testing history” for simulants entering the simulation, the above sampling strategy in effect assigns a prior BBBM test date to every eligile, low-propensity simulant when they are initialized. Namely, we can interpret the first draw \(W\) as the time the simulant waits to retest after a prior, negative BBBM test, and we can interpret the second draw \(T\) as the time it has been since that prior test. In this case the simulant’s prior BBBM test would have been at time \(t-T\), where \(t\) is the time the simulant enters the simulation. However, since we are not tracking this prior test date, we don’t know whether the simulant would have been eligible for testing at time \(t-T\) or whether their propensity would have been low enough to get a test at that time. Thus, the current strategy for selecting a next test date can’t be interpreted too literally as a “prior testing history,” and should instead be viewed mostly as a randomization strategy to avoid large numbers of simulants being tested immediately on initialization. On the other hand, if we do interpret this randomization strategy as assigning a testing history, note that we are assuming for simplicity that there were no prior false positive tests among simulants entering the simulation, so all previous BBBM tests are negative.
On timestep
On each timestep, simulants will have a chance to receive a BBBM test. This process should meet the following requirements:
Only simulants who are eligible based on the eligibility requirements for BBBM testing and whose propensity is below the time-specific testing rate can receive testing.
If a simulant meets both these criteria, check their next test date. If this date either corresponds to the current time step or is NaT, test the simulant now. (Thus, simulants who have not had a previous BBBM test will be tested as soon as they are eligible and the coverage rate increases above their propensity.)
For those who get tested, assign a positive diagnosis to 50% of people and a negative diagnosis to 50% of people. This 50% draw should be independent of any previous draws, e.g., people who test negative still have a 50% chance of being positive on a re-test.
If a simulant tests negative, the time of their next test is uniformly distributed between 3 and 5 years from the time of the negative test.
Record time of last test and yes/no diagnosis for determining future testing eligibility.
Note
Implementation: there are multiple possible ways to implement a uniform distribution for the waiting time between tests. The current model uses a “fortune-telling” implementation and assigns a future retest date between 3 and 5 years in the future.
Another mathematically equivalent option is to use a non-constant hazard function that increases the test liklihood on each timestep. This approach allows the decision of whether to test to be made “in the moment” rather than being predetermined, but it would require a new random draw on each time step. The discrete probability corresponding to this non-constant hazard function would be \(P(\text{test on the $k^\mathrm{th}$ time step}) = 1/(11 - k)\), for \(6\le k\le 10\). This results in a uniformly distributed population between 6 and 10 time steps, or 3 to 5 years. To conceptualize why this works, please see the table below outlining the time step value \(k\), the resulting probability of testing, and how a hypothetical population of 100 simulants is distributed over the time steps.
Time Step \(k\) |
Testing Probability |
People Tested |
Remaining Untested Population |
|---|---|---|---|
0-5 |
0% (ineligible) |
100 * 0% = 0 |
100 - 0 = 100 |
6 |
1/(11-6) = 20% |
100 * 20% = 20 |
100 - 20 = 80 |
7 |
1/(11-7) = 25% |
80 * 25% = 20 |
80 - 20 = 60 |
8 |
1/(11-8) = 33% |
60 * 33% = 20 |
60 - 20 = 40 |
9 |
1/(11-9) = 50% |
40 * 50% = 20 |
40 - 20 = 20 |
10 |
1/(11-10) = 100% |
20 * 100% = 20 |
20 - 20 = 0 |
Note
People who are not simulated (will not develop AD dementia) will also be tested. We counted these tests, including false positives, outside the simulation using a multistate life table (MSLT) model.
Assumptions and Limitations
Since BBBM testing eligibility is pre-clinical stage and CSF/PET is MCI or AD dementia stage, and simulants cannot move backwards, CSF/PET test history is irrelevant to BBBM test eligibility;
The same simulants undergo repeat testing to reflect ongoing issues with access or insurance, so propensity does not need to be re-assigned at any point;
Since BBBM uses the same propensity as existing testing, BBBM should replace many CSF and PET tests, though some simulants may not qualify for BBBM tests due to age requirements, or may get a BBBM false negative;
We determine whether a test is a false positive in the MSLT independently for healthy individuals, which precludes the possibility that individual characteristics like protein expression levels or chronic kidney disease drive heterogeneity in false positive rate;
We assume a false positive rate of zero among people who are simulated (will eventually develop AD dementia), which is inconsistent with our calculations in the MSLT; if the false positive rate were nonzero, some people would have prematurely started treatment before entering the simulation;
The strategy for implicitly assigning simulants’ BBBM test history does not account for the fact that simulants may not have been eligible for BBBM testing on all of the previous 10 time steps prior to entering the simulation; for example, we will implicitly assign a previous BBBM test date to a 65-year-old entering the simulation in, say, 2035 even though they wouldn’t have been eligible; the effects of this are hopefully small because improper testing can only happen during the first 5 years of the 20 years of eligible ages;
References
Roth S, Burnie N, Suridjan I, Yan JT, Carboni M. Current Diagnostic Pathways for Alzheimer’s Disease: A Cross-Sectional Real-World Study Across Six Countries. J Alzheimers Dis Rep. 7(1):659-674. doi:10.3233/ADR230007
Mattke S, Gustavsson A, Jacobs L, et al. Estimates of Current Capacity for Diagnosing Alzheimer’s Disease in Sweden and the Need to Expand Specialist Numbers. J Prev Alzheimers Dis. 2024;11(1):155-161. doi:10.14283/jpad.2023.94
Falahati F, Fereshtehnejad SM, Religa D, Wahlund LO, Westman E, Eriksdotter M. The Use of MRI, CT and Lumbar Puncture in Dementia Diagnostics: Data from the SveDem Registry. Dementia and Geriatric Cognitive Disorders. 2015;39(1-2):81-92. doi:10.1159/000366194